We present Segment Anything Model 2 (SAM 2), a foundation model towards solving promptable visual segmentation in images and videos. We build a data engine, which improves model and data via user interaction, to collect the largest video segmentation dataset to date. Our model is a simple transformer architecture with streaming memory for real-time video processing. SAM 2 trained on our data provides strong performance across a wide range of tasks. In video segmentation, we observe better accuracy, using 3x fewer interactions than prior approaches. In image segmentation, our model is more accurate and 6x faster than the Segment Anything Model (SAM). We believe that our data, model, and insights will serve as a significant milestone for video segmentation and related perception tasks. We are releasing a version of our model, the dataset and an interactive demo.

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Contrastive Language-Image Pre-training (CLIP) is an approach that has advanced research and applications in computer vision, fueling modern recognition systems and generative models. We believe that the main ingredient to the success of CLIP is its data and not the model architecture or pre-training objective. However, CLIP only provides very limited information about its data and how it has been collected, leading to works that aim to reproduce CLIP's data by filtering with its model parameters. In this work, we intend to reveal CLIP's data curation approach and in our pursuit of making it open to the community introduce Metadata-Curated Language-Image Pre-training (MetaCLIP). MetaCLIP takes a raw data pool and metadata (derived from CLIP's concepts) and yields a balanced subset over the metadata distribution. Our experimental study rigorously isolates the model and training settings, concentrating solely on data. MetaCLIP applied to CommonCrawl with 400M image-text data pairs outperforms CLIP's data on multiple standard benchmarks. In zero-shot ImageNet classification, MetaCLIP achieves 70.8% accuracy, surpassing CLIP's 68.3% on ViT-B models. Scaling to 1B data, while maintaining the same training budget, attains 72.4%. Our observations hold across various model sizes, exemplified by ViT-H achieving 80.5%, without any bells-and-whistles.

Window attention, position embeddings, and high resolution finetuning are core concepts in the modern transformer era of computer vision. However, we find that naively combining these near ubiquitous components can have a detrimental effect on performance. The issue is simple: interpolating position embeddings while using window attention is wrong. We study two state-of-the-art methods that have these three components, namely Hiera and ViTDet, and find that both do indeed suffer from this bug. To fix it, we introduce a simple absolute window position embedding strategy, which solves the bug outright in Hiera and allows us to increase both speed and performance of the model in ViTDet. We finally combine the two to obtain HieraDet, which achieves 61.7 box mAP on COCO, making it state-of-the-art for models that only use ImageNet-1k pretraining. This all stems from what is essentially a 3 line bug fix, which we name "absolute win".

Modern hierarchical vision transformers have added several vision-specific components in the pursuit of supervised classification performance. While these components lead to effective accuracies and attractive FLOP counts, the added complexity actually makes these transformers slower than their vanilla ViT counterparts. In this paper, we argue that this additional bulk is unnecessary. By pretraining with a strong visual pretext task (MAE), we can strip out all the bells-and-whistles from a state-of-the-art multi-stage vision transformer without losing accuracy. In the process, we create Hiera, an extremely simple hierarchical vision transformer that is more accurate than previous models while being significantly faster both at inference and during training. We evaluate Hiera on a variety of tasks for image and video recognition.

We present Masked Audio-Video Learners (MAViL) to train audio-visual representations. Our approach learns with three complementary forms of self-supervision: (1) reconstruction of masked audio and video input data, (2) intra- and inter-modal contrastive learning with masking, and (3) self-training by reconstructing joint audio-video contextualized features learned from the first two objectives. Pre-training with MAViL not only enables the model to perform well in audio-visual classification and retrieval tasks but also improves representations of each modality in isolation, without using information from the other modality for fine-tuning or inference. Empirically, MAViL sets a new state-of-the-art on AudioSet (53.1 mAP) and VGGSound (67.1% accuracy). For the first time, a self-supervised audio-visual model outperforms ones that use external supervision on these benchmarks.

Large vision-language models are generally applicable to many downstream tasks, but come at an exorbitant training cost that only large institutions can afford. This paper trades generality for efficiency and presents Curation in Training (CiT), a simple and efficient vision-text learning algorithm that couples a data objective into training. CiT automatically yields quality data to speed-up contrastive image-text training and alleviates the need for an offline data filtering pipeline, allowing broad data sources (including raw image-text pairs from the web). CiT contains two loops: an outer loop curating the training data and an inner loop consuming the curated training data. The text encoder connects the two loops. Given metadata for tasks of interest, e.g., class names, and a large pool of image-text pairs, CiT alternatively selects relevant training data from the pool by measuring the similarity of their text embeddings and embeddings of the metadata. In our experiments, we observe that CiT can speed up training by over an order of magnitude, especially if the raw data size is large.

This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional pre-pretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-pretraining scales with both model and data size making it applicable for training foundation models. Pre-pretraining consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images).

We present Fast Language-Image Pre-training (FLIP), a simple and more efficient method for training CLIP. Our method randomly masks out and removes a large portion of image patches during training. Masking allows us to learn from more image-text pairs given the same wall-clock time and contrast more samples per iteration with similar memory footprint. It leads to a favorable trade-off between accuracy and training time. In our experiments on 400 million image-text pairs, FLIP improves both accuracy and speed over the no-masking baseline. On a large diversity of downstream tasks, FLIP dominantly outperforms the CLIP counterparts trained on the same data. Facilitated by the speedup, we explore the scaling behavior of increasing the model size, data size, or training length, and report encouraging results and comparisons. We hope that our work will foster future research on scaling vision-language learning.

A central goal of visual recognition is to understand objects and scenes from a single image. 2D recognition has witnessed tremendous progress thanks to large-scale learning and general-purpose representations. Comparatively, 3D poses new challenges stemming from occlusions not depicted in the image. Prior works try to overcome these by inferring from multiple views or rely on scarce CAD models and category-specific priors which hinder scaling to novel settings. In this work, we explore single-view 3D reconstruction by learning generalizable representations inspired by advances in self-supervised learning. We introduce a simple framework that operates on 3D points of single objects or whole scenes coupled with category-agnostic large-scale training from diverse RGB-D videos. Our model, Multiview Compressive Coding (MCC), learns to compress the input appearance and geometry to predict the 3D structure by querying a 3D-aware decoder. MCC's generality and efficiency allow it to learn from large-scale and diverse data sources with strong generalization to novel objects imagined by DALL-E 2 or captured in-the-wild with an iPhone.

We introduce Token Merging (ToMe), a simple method to increase the throughput of existing ViT models without needing to train. ToMe gradually combines similar tokens in a transformer using a general and light-weight matching algorithm that is as fast as pruning while being more accurate. Off-the-shelf, ToMe can 2x the throughput of state-of-the-art ViT-L @ 512 and ViT-H @ 518 models on images and 2.2x the throughput of ViT-L on video with only a 0.2-0.3% accuracy drop in each case. ToMe can also easily be applied during training, improving in practice training speed up to 2x for MAE fine-tuning on video. Training with ToMe further minimizes accuracy drop, leading to 2x the throughput of ViT-B on audio for only a 0.4% mAP drop. Qualitatively, we find that ToMe merges object parts into one token, even over multiple frames of video. Overall, ToMe's accuracy and speed are competitive with state-of-the-art on images, video, and audio.

This paper studies a conceptually simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos. We randomly mask out spacetime patches in videos and learn an autoencoder to reconstruct them in pixels. Interestingly, we show that our MAE method can learn strong representations with almost no inductive bias on spacetime (only except for patch and positional embeddings), and spacetime-agnostic random masking performs the best. We observe that the optimal masking ratio is as high as 90% (vs. 75% on images), supporting the hypothesis that this ratio is related to information redundancy of the data. A high masking ratio leads to a large speedup, e.g., > 4x in wall-clock time or even more. We report competitive results on several challenging video datasets using vanilla Vision Transformers. We observe that MAE can outperform supervised pre-training by large margins. We further report encouraging results of training on real-world, uncurated Instagram data. Our study suggests that the general framework of masked autoencoding (BERT, MAE, etc.) can be a unified methodology for representation learning with minimal domain knowledge.

This paper studies a simple extension of image-based Masked Autoencoders (MAE) to self-supervised representation learning from audio spectrograms. Following the Transformer encoder-decoder design in MAE, our Audio-MAE first encodes audio spectrogram patches with a high masking ratio, feeding only the non-masked tokens through encoder layers. The decoder then re-orders and decodes the encoded context padded with mask tokens, in order to reconstruct the input spectrogram. We find it beneficial to incorporate local window attention in the decoder, as audio spectrograms are highly correlated in local time and frequency bands. We then fine-tune the encoder with a lower masking ratio on target datasets. Empirically, Audio-MAE sets new state-of-the-art performance on six audio and speech classification tasks, outperforming other recent models that use external supervised pre-training.

In this paper, we propose a new strategy to overcome this challenge. Instead of trying to process more frames at once like most existing methods, we propose to process videos in an online fashion and cache "memory" at each iteration. Through the memory, the model can reference prior context for long-term modeling, with only a marginal cost. Based on this idea, we build MeMViT, a Memory-augmented Multiscale Vision Transformer, that has a temporal support 30x longer than existing models with only 4.5% more compute; traditional methods need >3,000% more compute to do the same. On a wide range of settings, the increased temporal support enabled by MeMViT brings large gains in recognition accuracy consistently. MeMViT obtains state-of-the-art results on the AVA, EPIC-Kitchens-100 action classification, and action anticipation datasets. Code and models will be made publicly available.

We present Masked Feature Prediction (MaskFeat) for self-supervised pre-training of video models. Our approach first randomly masks out a portion of the input sequence and then predicts the feature of the masked regions. We study five different types of features and find Histograms of Oriented Gradients (HOG), a hand-crafted feature descriptor, works particularly well in terms of both performance and efficiency. We observe that the local contrast normalization in HOG is essential for good results, which is in line with earlier work using HOG for visual recognition. Our approach can learn abundant visual knowledge and drive large-scale Transformer-based models. Without using extra model weights or supervision, MaskFeat pre-trained on unlabeled videos achieves unprecedented results of 86.7% with MViT-L on Kinetics-400, 88.3% on Kinetics-600, 80.4% on Kinetics-700, 38.8 mAP on AVA, and 75.0% on SSv2. MaskFeat further generalizes to image input, which can be interpreted as a video with a single frame and obtains competitive results on ImageNet.

In this paper, we study Multiscale Vision Transformers (MViT) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTs' pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViT has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 56.1 box AP on COCO object detection as well as 86.1% on Kinetics-400 video classification.

We present Reversible Vision Transformers, a memory efficient architecture design for visual recognition. By decoupling the GPU memory footprint from the depth of the model, Reversible Vision Transformers enable memory efficient scaling of transformer architectures. We adapt two popular models, namely Vision Transformer and Multi-scale Vision Transformers, to reversible variants and benchmark extensively across both model sizes and tasks of image classification, object detection and video classification. Reversible Vision Transformers achieve a reduced memory footprint of up to 15.5x at identical model complexity, parameters and accuracy, demonstrating the promise of reversible vision transformers as an efficient backbone for resource limited training regimes. Finally, we find that the additional computational burden of recomputing activations is more than overcome for deeper models, where throughput can increase up to 3.9x over their non-reversible counterparts.

The "Roaring 20s" of visual recognition began with the introduction of Vision Transformers (ViTs), which quickly superseded ConvNets as the state-of-the-art image classification model. A vanilla ViT, on the other hand, faces difficulties when applied to general computer vision tasks such as object detection and semantic segmentation. It is the hierarchical Transformers (e.g., Swin Transformers) that reintroduced several ConvNet priors, making Transformers practically viable as a generic vision backbone and demonstrating remarkable performance on a wide variety of vision tasks. However, the effectiveness of such hybrid approaches is still largely credited to the intrinsic superiority of Transformers, rather than the inherent inductive biases of convolutions. In this work, we reexamine the design spaces and test the limits of what a pure ConvNet can achieve. We gradually "modernize" a standard ResNet toward the design of a vision Transformer, and discover several key components that contribute to the performance difference along the way. The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt. Constructed entirely from standard ConvNet modules, ConvNeXts compete favorably with Transformers in terms of accuracy and scalability, achieving 87.8% ImageNet top-1 accuracy and outperforming Swin Transformers on COCO detection and ADE20K segmentation, while maintaining the simplicity and efficiency of standard ConvNets.

We present TrackFormer, an end-to-end multi-object tracking and segmentation model based on an encoder-decoder Transformer architecture. Our approach introduces track query embeddings which follow objects through a video sequence in an autoregressive fashion. New track queries are spawned by the DETR object detector and embed the position of their corresponding object over time. The Transformer decoder adjusts track query embeddings from frame to frame, thereby following the changing object positions. TrackFormer achieves a seamless data association between frames in a new tracking-by-attention paradigm by self- and encoder-decoder attention mechanisms which simultaneously reason about location, occlusion, and object identity. TrackFormer yields state-of-the-art performance on the tasks of multi-object tracking (MOT17) and segmentation (MOTS20). We hope our unified way of performing detection and tracking will foster future research in multi-object tracking and video understanding.

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,025 hours of daily-life activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 855 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception.

We introduce PyTorchVideo, an open-source deep-learning library that provides a rich set of modular, efficient, and reproducible components for a variety of video understanding tasks, including classification, detection, self-supervised learning, and low-level processing. The library covers a full stack of video understanding tools including multimodal data loading, transformations, and models that reproduce state-of-the-art performance. PyTorchVideo further supports hardware acceleration that enables real-time inference on mobile devices. The library is based on PyTorch and can be used by any training framework; for example, PyTorchLightning, PySlowFast, or Classy Vision.

We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10x more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers.

We present a large-scale study on unsupervised spatiotemporal representation learning from videos. With a unified perspective on four recent image-based frameworks, we study a simple objective that can easily generalize all these methods to space-time. Our objective encourages temporally-persistent features in the same video, and in spite of its simplicity, it works surprisingly well across: (i) different unsupervised frameworks, (ii) pre-training datasets, (iii) downstream datasets, and (iv) backbone architectures. We draw a series of intriguing observations from this study, e.g., we discover that encouraging long-spanned persistency can be effective even if the timespan is 60 seconds. In addition to state-of-the-art results in multiple benchmarks, we report a few promising cases in which unsupervised pre-training can outperform its supervised counterpart.

We present a multiview pseudo-labeling approach to video learning, a novel framework that uses complementary views in the form of appearance and motion information for semi-supervised learning in video. The complementary views help obtain more reliable pseudo-labels on unlabeled video, to learn stronger video representations than from purely supervised data. Though our method capitalizes on multiple views, it nonetheless trains a model that is shared across appearance and motion input and thus, by design, incurs no additional computation overhead at inference time. On multiple video recognition datasets, our method substantially outperforms its supervised counterpart, and compares favorably to previous work on standard benchmarks in self-supervised video representation learning.

We present VideoCLIP, a contrastive approach to pre-train a unified model for zero-shot video and text understanding, without using any labels on downstream tasks. VideoCLIP trains a transformer for video and text by contrasting temporally overlapping positive video-text pairs with hard negatives from nearest neighbor retrieval. Our experiments on a diverse series of downstream tasks, including sequence-level text-video retrieval, VideoQA, token-level action localization, and action segmentation reveal state-of-the-art performance, surpassing prior work, and in some cases even outperforming supervised approaches.

In video transformers, the time dimension is often treated in the same way as the two spatial dimensions. However, in a scene where objects or the camera may move, a physical point imaged at one location in frame may be entirely unrelated to what is found at that location in frame . These temporal correspondences should be modeled to facilitate learning about dynamic scenes. To this end, we propose a new drop-in block for video transformers -- trajectory attention -- that aggregates information along implicitly determined motion paths. We additionally propose a new method to address the quadratic dependence of computation and memory on the input size, which is particularly important for high resolution or long videos. While these ideas are useful in a range of settings, we apply them to the specific task of video action recognition with a transformer model and obtain state-of-the-art results on the Kinetics, Something--Something V2, and Epic-Kitchens datasets.

We present a simplified, task-agnostic multi-modal pre-training approach that can accept either video or text input, or both for a variety of end tasks. Existing pre-training are task-specific by adopting either a single cross-modal encoder that requires both modalities, limiting their use for retrieval-style end tasks or more complex multitask learning with two unimodal encoders, limiting early cross-modal fusion. We instead introduce new pretraining masking schemes that better mix across modalities (e.g. by forcing masks for text to predict the closest video embeddings) while also maintaining separability (e.g. unimodal predictions are sometimes required, without using all the input). Experimental results show strong performance across a wider range of tasks than any previous methods, often outperforming task-specific pre-training

This paper presents X3D, a family of efficient video networks that progressively expand a tiny 2D image classification architecture along multiple network axes, in space, time, width and depth. Inspired by feature selection methods in machine learning, a simple stepwise network expansion approach is employed that expands a single axis in each step, such that good accuracy to complexity trade-off is achieved. To expand X3D to a specific target complexity, we perform progressive forward expansion followed by backward contraction. X3D achieves state-of-the-art performance while requiring 4.8x and 5.5x fewer multiply-adds and parameters for similar accuracy as previous work. Our most surprising finding is that networks with high spatiotemporal resolution can perform well, while being extremely light in terms of network width and parameters. We report competitive accuracy at unprecedented efficiency on video classification and detection benchmarks.

We present Audiovisual SlowFast Networks, an architecture for integrated audiovisual perception. AVSlowFast has Slow and Fast visual pathways that are deeply integrated with a Faster Audio pathway to model vision and sound in a unified representation. We fuse audio and visual features at multiple layers, enabling audio to contribute to the formation of hierarchical audiovisual concepts. To overcome training difficulties that arise from different learning dynamics for audio and visual modalities, we introduce DropPathway, which randomly drops the Audio pathway during training as an effective regularization technique. Inspired by prior studies in neuroscience, we perform hierarchical audiovisual synchronization to learn joint audiovisual features. We report state-of-the-art results on six video action classification and detection datasets, perform detailed ablation studies, and show the generalization of AVSlowFast to learn self-supervised audiovisual features.

Feature pyramid networks (FPN) have been widely adopted in the object detection literature to improve feature representations for better handling of variations in scale. In this paper, we present Feature Pyramid Grids (FPG), a simple extension to FPN, that represents the feature scale-space as a regular grid of parallel bottom-up pathways which are fused by multi-directional lateral connections between them. FPG is simple and flexible, which only adds a small overhead to regular, single pathway FPN while significantly increasing its performance. In addition to its general and simple structure, over complicated structures that have been found with neural architecture search, it also compares favorably against such approaches, providing higher accuracy and speed. We hope that FPG with its simple and effective nature can serve as a strong baseline for future work in object recognition.

Training competitive deep video models is an order of magnitude slower than training their counterpart image models. Slow training causes long research cycles, which hinders progress in video understanding research. Following standard practice for training image models, video model training assumes a fixed mini-batch shape: a specific number of clips, frames, and spatial size. However, what is the optimal shape? High resolution models perform well, but train slowly. Low resolution models train faster, but they are inaccurate. Inspired by multigrid methods in numerical optimization, we propose to use variable mini-batch shapes with different spatial-temporal resolutions that are varied according to a schedule. The different shapes arise from resampling the training data on multiple sampling grids. Training is accelerated by scaling up the mini-batch size and learning rate when shrinking the other dimensions. We empirically demonstrate a general and robust grid schedule that yields a significant out-of-the-box training speedup without a loss in accuracy for different models (I3D, non-local, SlowFast), datasets (Kinetics, Something-Something, Charades), and training settings (with and without pre-training, 128 GPUs or 1 GPU). As an illustrative example, the proposed multigrid method trains a ResNet-50 SlowFast network 4.5x faster (wall-clock time, same hardware) while also improving accuracy (+0.8% absolute) on Kinetics-400 compared to the baseline training method.

First-person video naturally brings the use of a physical environment to the forefront, since it shows the camera wearer interacting fluidly in a space based on his intentions. However, current methods largely separate the observed actions from the persistent space itself. We introduce a model for environment affordances that is learned directly from egocentric video. The main idea is to gain a human-centric model of a physical space (such as a kitchen) that captures (1) the primary spatial zones of interaction and (2) the likely activities they support. Our approach decomposes a space into a topological map derived from first-person activity, organizing an ego-video into a series of visits to the different zones. Further, we show how to link zones across multiple related environments (e.g., from videos of multiple kitchens) to obtain a consolidated representation of environment functionality. On EPIC-Kitchens and EGTEA+, we demonstrate our approach for learning scene affordances and anticipating future actions in long-form video.

We present SlowFast networks for video recognition. Our model involves (i) a Slow pathway, operating at low frame rate, to capture spatial semantics, and (ii) a Fast pathway, operating at high frame rate, to capture motion at fine temporal resolution. The Fast pathway can be made very lightweight by reducing its channel capacity, yet can learn useful temporal information for video recognition. Our models achieve strong performance for both action classification and detection in video, and large improvements are pin-pointed as contributions by our SlowFast concept. We report 79.0% accuracy on the Kinetics dataset without using any pre-training, largely surpassing the previous best results of this kind. On AVA action detection we achieve a new state-of-the-art of 28.3 mAP.

Previous work on predicting or generating 3D human pose sequences regresses either joint rotations or joint positions. The former strategy is prone to error accumulation along the kinematic chain, as well as discontinuities when using Euler angles or exponential maps as parameterizations. The latter requires re-projection onto skeleton constraints to avoid bone stretching and invalid configurations. This work addresses both limitations. QuaterNet represents rotations with quaternions and our loss function performs forward kinematics on a skeleton to penalize absolute position errors instead of angle errors. We investigate both recurrent and convolutional architectures and evaluate on short-term prediction and long-term generation. For the latter, our approach is qualitatively judged as realistic as recent neural strategies from the graphics literature. Our experiments compare quaternions to Euler angles as well as exponential maps and show that only a very short context is required to make reliable future predictions. Finally, we show that the standard evaluation protocol for Human3.6M produces high variance results and we propose a simple solution.

Learning how to interact with objects is an important step towards embodied visual intelligence, but existing techniques suffer from heavy supervision or sensing requirements. We propose an approach to learn human-object interaction "hotspots" directly from video. Rather than treat affordances as a manually supervised semantic segmentation task, our approach learns about interactions by watching videos of real human behavior and anticipating afforded actions. Given a novel image or video, our model infers a spatial hotspot map indicating how an object would be manipulated in a potential interaction-- even if the object is currently at rest. Through results with both first and third person video, we show the value of grounding affordances in real human-object interactions. Not only are our weakly supervised hotspots competitive with strongly supervised affordance methods, but they can also anticipate object interaction for novel object categories.

Modern approaches for multi-person pose estimation in video require large amounts of dense annotations. However, labeling every frame in a video is costly and labor intensive. To reduce the need for dense annotations, we propose a PoseWarper network that leverages training videos with sparse annotations (every k frames) to learn to perform dense temporal pose propagation and estimation. Given a pair of video frames---a labeled Frame A and an unlabeled Frame B---we train our model to predict human pose in Frame A using the features from Frame B by means of deformable convolutions to implicitly learn the pose warping between A and B. We demonstrate that we can leverage our trained PoseWarper for several applications. First, at inference time we can reverse the application direction of our network in order to propagate pose information from manually annotated frames to unlabeled frames. This makes it possible to generate pose annotations for the entire video given only a few manually-labeled frames. Compared to modern label propagation methods based on optical flow, our warping mechanism is much more compact (6M vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also show that we can improve the accuracy of a pose estimator by training it on an augmented dataset obtained by adding our propagated poses to the original manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose information from neighboring frames during inference. This allows our system to achieve state-of-the-art pose detection results on the PoseTrack2017 dataset.

To understand the world, we humans constantly need to relate the present to the past, and put events in context. In this paper, we enable existing video models to do the same. We propose a long-term feature bank---supportive information extracted over the entire span of a video---to augment state-of-the-art video models that otherwise would only view short clips of 2-5 seconds. Our experiments demonstrate that augmenting 3D convolutional networks with a long-term feature bank yields state-of-the-art results on three challenging video datasets: AVA, EPIC-Kitchens, and Charades.

In this work, we demonstrate that 3D poses in video can be effectively estimated with a fully convolutional model based on dilated temporal convolutions over 2D keypoints. We also introduce back-projection, a simple and effective semi-supervised training method that leverages unlabeled video data. We start with predicted 2D keypoints for unlabeled video, then estimate 3D poses and finally back-project to the input 2D keypoints. In the supervised setting, our fully-convolutional model outperforms the previous best result from the literature by 6 mm mean per-joint position error on Human3.6M, corresponding to an error reduction of 11%, and the model also shows significant improvements on HumanEva-I. Moreover, experiments with back-projection show that it comfortably outperforms previous state-of-the-art results in semi-supervised settings where labeled data is scarce.

As the success of deep models has led to their deployment in all areas of computer vision, it is increasingly important to understand how these representations work and what they are capturing. In this paper, we shed light on deep spatiotemporal representations by visualizing what two-stream models have learned in order to recognize actions in video. We show that local detectors for appearance and motion objects arise to form distributed representations for recognizing human actions. Key observations include the following. First, cross-stream fusion enables the learning of true spatiotemporal features rather than simply separate appearance and motion features. Second, the networks can learn local representations that are highly class specific, but also generic representations that can serve a range of classes. Third, throughout the hierarchy of the network, features become more abstract and show increasing invariance to aspects of the data that are unimportant to desired distinctions (e.g. motion patterns across various speeds). Fourth, visualizations can be used not only to shed light on learned representations, but also to reveal idiosyncracies of training data and to explain failure cases of the system.

Despite huge success in the image domain, modern detection models such as Faster R-CNN have not been used nearly as much for video analysis. This is arguably due to the fact that detection models are designed to operate on single frames and as a result do not have a mechanism for learning motion representations directly from video. We propose a learning procedure that allows detection models such as Faster R-CNN to learn motion features directly from the RGB video data while being optimized with respect to a pose estimation task. In our experiments we show that our training scheme helps learn effective motion cues, which can be used to estimate and localize salient human motion. Furthermore, we demonstrate that as a byproduct, our model also learns features that lead to improved pose detection in still-images, and better keypoint tracking. Finally, we show how to leverage our learned model for the tasks of spatiotemporal action localization and fine-grained action recognition.

This paper documents the winning entry at the CVPR2017 vehicle velocity estimation challenge. Velocity estimation is an emerging task in autonomous driving which has not yet been thoroughly explored. The goal is to estimate the relative velocity of a specific vehicle from a sequence of images. In this paper, we present a light-weight approach for directly regressing vehicle velocities from their trajectories using a multilayer perceptron. Another contribution is an explorative study of features for monocular vehicle velocity estimation. We find that lightweight trajectory based features outperform depth and motion cues extracted from deep ConvNets, especially for far-distance predictions where current disparity and optical flow estimators are challenged significantly. Our light-weight approach is real-time capable on a single CPU and outperforms all competing entries in the velocity estimation challenge. On the test set, we report an average error of 1.12 m/s which is comparable to a (ground-truth) system that combines LiDAR and radar techniques to achieve an error of around 0.71 m/s.

Recent approaches for high accuracy detection and tracking of object categories in video consist of complex multistage solutions that become more cumbersome each year. In this paper we propose a ConvNet architecture that jointly performs detection and tracking, solving the task in a simple and effective way. Our contributions are threefold: (i) we set up a ConvNet architecture for simultaneous detection and tracking, using a multi-task objective for frame-based object detection and across-frame track regression; (ii) we introduce novel correlation features that represent object co-occurrences across time to aid the ConvNet during tracking; (iii) we link the frame level detections based on our across-frame tracklets to produce high accuracy detections at the video level. Our ConvNet architecture for spatiotemporal object detection is evaluated on the large-scale ImageNet VID dataset where it achieves state-of-the-art results. Our approach provides better single model performance than the winning method of the last ImageNet challenge while being conceptually much simpler. Finally, we show that by increasing the temporal stride we can dramatically increase the tracker speed.

This paper presents a general ConvNet architecture for video action recognition based on multiplicative interactions of spacetime features. Our model combines the appearance and motion pathways of a two-stream architecture by motion gating and is trained end-to-end. We theoretically motivate multiplicative gating functions for residual networks and empirically study their effect on classification accuracy. To capture long-term dependencies we inject identity mapping kernels for learning temporal relationships. Our architecture is fully convolutional in spacetime and able to evaluate a video in a single forward pass. Empirical investigation reveals that our model produces state-of-the-art results on two standard action recognition datasets.

This paper combines three contributions to establish a new state-of-the-art in dynamic scene recognition. First, we present a novel ConvNet architecture based on temporal residual units that is fully convolutional in spacetime. Our model augments spatial ResNets with convolutions across time to hierarchically add temporal residuals as the depth of the network increases. Second, existing approaches to video-based recognition are categorized and a baseline of seven previously top performing algorithms is selected for comparative evaluation on dynamic scenes. Third, we introduce a new and challenging video database of dynamic scenes that more than doubles the size of those previously available.This dataset is explicitly split into two subsets of equal size that contain videos with and without camera motion to allow for systematic study of how this variable interacts with the defining dynamics of the scene per se. Our evaluations verify the particular strengths and weaknesses of the baseline algorithms with respect to various scene classes and camera motion parameters. Finally, our temporal ResNet boosts recognition performance and establishes a new state-of-the-art on dynamic scene recognition, as well as on the complementary task of action recognition.

Two-stream Convolutional Networks (ConvNets) have shown strong performance for human action recognition in videos. Recently, Residual Networks (ResNets) have arisen as a new technique to train extremely deep architectures. In this paper, we introduce spatiotemporal ResNets as a combination of these two approaches. Our novel architecture generalizes ResNets for the spatiotemporal domain by introducing residual connections in two ways. First, we inject residual connections between the appearance and motion pathways of a two-stream architecture to allow spatiotemporal interaction between the two streams. Second, we transform pretrained image ConvNets into spatiotemporal networks by equipping these with learnable convolutional filters that are initialized as temporal residual connections and operate on adjacent feature maps in time. This approach slowly increases the spatiotemporal receptive field as the depth of the model increases and naturally integrates image ConvNet design principles. The whole model is trained end-to-end to allow hierarchical learning of complex spatiotemporal features. We evaluate our novel spatiotemporal ResNet using two widely used action recognition benchmarks where it exceeds the previous state-of-the-art.

Recent applications of Convolutional Neural Networks (ConvNets) for human action recognition in videos have proposed different solutions for incorporating the appearance and motion information. We study a number of ways of fusing ConvNet towers both spatially and temporally in order to best take advantage of this spatio-temporal information. We make the following findings: (i) that rather than fusing at the softmax layer, a spatial and temporal network can be fused at a convolution layer without loss of performance, but with a substantial saving in parameters; (ii) that it is better to fuse such networks spatially at the last convolutional layer than earlier, and that additionally fusing at the class prediction layer can boost accuracy; finally (iii) that pooling of abstract convolutional features over spatiotemporal neighbourhoods further boosts performance. Based on these studies we propose a new ConvNet architecture for spatiotemporal fusion of video snippets, and evaluate its performance on standard benchmarks where this architecture achieves state-of-the-art results.

This paper presents Dynamically Pooled Complementary Features, a unified approach to dynamic scene recognition that analyzes a short video clip in terms of its spatial, temporal and color properties. The complementarity of these properties is preserved through all main steps of processing, including primitive feature extraction, coding and pooling. In the feature extraction step, spatial orientations capture static appearance, spatiotemporal oriented energies capture image dynamics and color statistics capture chromatic information. Subsequently, primitive features are encoded into a mid-level representation that has been learned for the task of dynamic scene recognition. Finally, a novel dynamic spacetime pyramid is introduced. This dynamic pooling approach can handle both global as well as local motion by adapting to the temporal structure, as guided by pooling energies. The resulting system provides online recognition of dynamic scenes that is thoroughly evaluated on the two current benchmark datasets and yields best results to date on both datasets. In-depth analysis reveals the benefits of explicitly modeling feature complementarity in combination with the dynamic spacetime pyramid, indicating that this unified approach should be well-suited to many areas of video analysis.

Human actions typically occur over a well localized extent in both space and time. Similarly, as typically captured in video, human actions have small spatiotemporal support in image space. This paper capitalizes on these observations by weighting feature pooling for action recognition over those areas within a video where actions are most likely to occur. To enable this operation, we define a novel measure of spacetime saliency. The measure relies on two observations regarding foreground motion of human actors: They typically exhibit motion that contrasts with that of their surrounding region and they are spatially compact. By using the resulting definition of saliency during feature pooling we show that action recognition performance achieves state-of-the-art levels on three widely considered action recognition datasets. Our saliency weighted pooling can be applied to essentially any locally defined features and encodings thereof. Additionally, we demonstrate that inclusion of locally aggregated spatiotemporal energy features, which efficiently result as a by-product of the saliency computation, further boosts performance over reliance on standard action recognition features alone.

This paper presents a unified bag of visual word (BoW) framework for dynamic scene recognition. The approach builds on primitive features that uniformly capture spatial and temporal orientation structure of the imagery (e.g., video), as extracted via application of a bank of spatiotemporally oriented filters. Various feature encoding techniques are investigated to abstract the primitives to an intermediate representation that is best suited to dynamic scene representation. Further, a novel approach to adaptive pooling of the encoded features is presented that captures spatial layout of the scene even while being robust to situations where camera motion and scene dynamics are confounded. The resulting overall approach has been evaluated on two standard, publically available dynamic scene datasets. The results show that in comparison to a representative set of alternatives, the proposed approach outperforms the previous state-of-the-art in classification accuracy by 10%.

The combination of RFID and computer vision systems is an effective approach to mitigate the limited tag localization capabilities of current RFID deployments. In this paper, we present a hybrid RFID and computer vision system for localization and tracking of RFID tags. The proposed system combines the information from the two complementary sensor modalities in a probabilistic manner and provides a high degree of flexibility. In addition, we introduce a robust data association method which is crucial for the application in practical scenarios. To demonstrate the performance of the proposed system, we conduct a series of experiments in an article surveillance setup. This is a frequent application for RFID systems in retail where previous approaches solely based on RFID localization have difficulties due to false alarms triggered by stationary tags. Our evaluation shows that the fusion of RFID and computer vision provides robustness to false positive observations and allows for a reliable system operation.

This paper presents spacetime forests defined over complementary spatial and temporal features for recognition of naturally occurring dynamic scenes. The approach improves on the previous state-of-the-art in both classification and execution rates. A particular improvement is with increased robustness to camera motion, where previous approaches have experienced difficulty. There are three key novelties in the approach. First, a novel spacetime descriptor is employed that exploits the complementary nature of spatial and temporal information, as inspired by previous research on the role of orientation features in scene classification. Second, a forest-based classifier is used to learn a multi-class representation of the feature distributions. Third, the video is processed in temporal slices with scale matched preferentially to scene dynamics over camera motion. Slicing allows for temporal alignment to be handled as latent information in the classifier and for efficient, incremental processing. The integrated approach is evaluated empirically on two publically available datasets to document its outstanding performance.

This paper presents two fundamental contributions that can be very useful for any autonomous system that requires point correspondences for visual odometry. First, the Spatio-Temporal Monitor (STM) is an efficient method to identify good features to track by monitoring their spatiotemporal (x-y-t) appearance without any assumptions about motion or geometry. The STM may be used with any spatial (x-y) descriptor, but it performs best when combined with our second contribution, the Histogram of Oriented Magnitudes (HOM) descriptor, which is based on spatially oriented multiscale filter magnitudes. To fulfil the real-time requirements of autonomous applications, the same descriptor can be used for both, track generation and monitoring, to identify low-quality feature tracks at virtually no additional computational cost. Our extensive experimental validation on a challenging public dataset demonstrates the excellent performance of STM and HOM, where we significantly outperform the well known “Good Features to Track” method and show that our proposed feature quality measure highly correlates with the accuracy in structure and motion estimation.

In this letter, a no-reference perceptual sharpness metric based on a statistical analysis of local edge gradients is presented. The method takes properties of the human visual system into account. Based on perceptual properties, a relationship between the extracted statistical features and the metric score is established to form a Perceptual Sharpness Index (PSI). A comparison with state-of-the-art metrics shows that the proposed method correlates highly with human perception and exhibits low computational complexity. In contrast to existing metrics, the PSI performs well for a wide range of blurriness and shows a high degree of invariance for different image contents.