It is challenging for artificial intelligence systems to achieve accurate video recognition under the scenario of low computation costs. Adaptive inference based efficient video recognition methods typically preview videos and focus on salient parts to reduce computation costs. Most existing works focus on complex networks learning with video classification based objectives. Taking all frames as positive samples, few of them pay attention to the discrimination between positive samples (salient frames) and negative samples (non-salient frames) in supervisions. To fill this gap, in this paper, we propose a novel Non-saliency Suppression Network (NSNet), which effectively suppresses the responses of non-salient frames. Specifically, on the frame level, effective pseudo labels that can distinguish between salient and non-salient frames are generated to guide the frame saliency learning. On the video level, a temporal attention module is learned under dual video-level supervisions on both the salient and the non-salient representations. Saliency measurements from both two levels are combined for exploitation of multi-granularity complementary information. Extensive experiments conducted on four well-known benchmarks verify our NSNet not only achieves the state-of-the-art accuracy-efficiency trade-off but also present a significantly faster (2.4~4.3x) practical inference speed than state-of-the-art methods. Our project page is at https://lawrencexia2008.github.io/projects/nsnet .
Clustering is a ubiquitous tool in unsupervised learning. Most of the existing self-supervised representation learning methods typically cluster samples based on visually dominant features. While this works well for image-based self-supervision, it often fails for videos, which require understanding motion rather than focusing on background. Using optical flow as complementary information to RGB can alleviate this problem. However, we observe that a naive combination of the two views does not provide meaningful gains. In this paper, we propose a principled way to combine two views. Specifically, we propose a novel clustering strategy where we use the initial cluster assignment of each view as prior to guide the final cluster assignment of the other view. This idea will enforce similar cluster structures for both views, and the formed clusters will be semantically abstract and robust to noisy inputs coming from each individual view. Additionally, we propose a novel regularization strategy to address the feature collapse problem, which is common in cluster-based self-supervised learning methods. Our extensive evaluation shows the effectiveness of our learned representations on downstream tasks, e.g., video retrieval and action recognition. Specifically, we outperform the state of the art by 7% on UCF and 4% on HMDB for video retrieval, and 5% on UCF and 6% on HMDB for video classification
We present a new benchmark dataset, Sapsucker Woods 60 (SSW60), for advancing research on audiovisual fine-grained categorization. While our community has made great strides in fine-grained visual categorization on images, the counterparts in audio and video fine-grained categorization are relatively unexplored. To encourage advancements in this space, we have carefully constructed the SSW60 dataset to enable researchers to experiment with classifying the same set of categories in three different modalities: images, audio, and video. The dataset covers 60 species of birds and is comprised of images from existing datasets, and brand new, expert-curated audio and video datasets. We thoroughly benchmark audiovisual classification performance and modality fusion experiments through the use of state-of-the-art transformer methods. Our findings show that performance of audiovisual fusion methods is better than using exclusively image or audio based methods for the task of video classification. We also present interesting modality transfer experiments, enabled by the unique construction of SSW60 to encompass three different modalities. We hope the SSW60 dataset and accompanying baselines spur research in this fascinating area.
Audio-visual generalised zero-shot learning for video classification requires understanding the relations between the audio and visual information in order to be able to recognise samples from novel, previously unseen classes at test time. The natural semantic and temporal alignment between audio and visual data in video data can be exploited to learn powerful representations that generalise to unseen classes at test time. We propose a multi-modal and Temporal Cross-attention Framework (\modelName) for audio-visual generalised zero-shot learning. Its inputs are temporally aligned audio and visual features that are obtained from pre-trained networks. Encouraging the framework to focus on cross-modal correspondence across time instead of self-attention within the modalities boosts the performance significantly. We show that our proposed framework that ingests temporal features yields state-of-the-art performance on the \ucf, \vgg, and \activity benchmarks for (generalised) zero-shot learning. Code for reproducing all results is available at \url{https://github.com/ExplainableML/TCAF-GZSL}.
Utilizing vision and language models (VLMs) pre-trained on large-scale image-text pairs is becoming a promising paradigm for open-vocabulary visual recognition. In this work, we extend this paradigm by leveraging motion and audio that naturally exist in video. We present \textbf{MOV}, a simple yet effective method for \textbf{M}ultimodal \textbf{O}pen-\textbf{V}ocabulary video classification. In MOV, we directly use the vision encoder from pre-trained VLMs with minimal modifications to encode video, optical flow and audio spectrogram. We design a cross-modal fusion mechanism to aggregate complimentary multimodal information. Experiments on Kinetics-700 and VGGSound show that introducing flow or audio modality brings large performance gains over the pre-trained VLM and existing methods. Specifically, MOV greatly improves the accuracy on base classes, while generalizes better on novel classes. MOV achieves state-of-the-art results on UCF and HMDB zero-shot video classification benchmarks, significantly outperforming both traditional zero-shot methods and recent methods based on VLMs. Code and models will be released.
Video transformer naturally incurs a heavier computation burden than a static vision transformer, as the former processes $T$ times longer sequence than the latter under the current attention of quadratic complexity $(T^2N^2)$. The existing works treat the temporal axis as a simple extension of spatial axes, focusing on shortening the spatio-temporal sequence by either generic pooling or local windowing without utilizing temporal redundancy. However, videos naturally contain redundant information between neighboring frames; thereby, we could potentially suppress attention on visually similar frames in a dilated manner. Based on this hypothesis, we propose the LAPS, a long-term ``\textbf{\textit{Leap Attention}}'' (LA), short-term ``\textbf{\textit{Periodic Shift}}'' (\textit{P}-Shift) module for video transformers, with $(2TN^2)$ complexity. Specifically, the ``LA'' groups long-term frames into pairs, then refactors each discrete pair via attention. The ``\textit{P}-Shift'' exchanges features between temporal neighbors to confront the loss of short-term dynamics. By replacing a vanilla 2D attention with the LAPS, we could adapt a static transformer into a video one, with zero extra parameters and neglectable computation overhead ($\sim$2.6\%). Experiments on the standard Kinetics-400 benchmark demonstrate that our LAPS transformer could achieve competitive performances in terms of accuracy, FLOPs, and Params among CNN and transformer SOTAs. We open-source our project in \sloppy \href{https://github.com/VideoNetworks/LAPS-transformer}{\textit{\color{magenta}{https://github.com/VideoNetworks/LAPS-transformer}}} .
Transfer learning is the predominant paradigm for training deep networks on small target datasets. Models are typically pretrained on large ``upstream'' datasets for classification, as such labels are easy to collect, and then finetuned on ``downstream'' tasks such as action localisation, which are smaller due to their finer-grained annotations. In this paper, we question this approach, and propose co-finetuning -- simultaneously training a single model on multiple ``upstream'' and ``downstream'' tasks. We demonstrate that co-finetuning outperforms traditional transfer learning when using the same total amount of data, and also show how we can easily extend our approach to multiple ``upstream'' datasets to further improve performance. In particular, co-finetuning significantly improves the performance on rare classes in our downstream task, as it has a regularising effect, and enables the network to learn feature representations that transfer between different datasets. Finally, we observe how co-finetuning with public, video classification datasets, we are able to achieve state-of-the-art results for spatio-temporal action localisation on the challenging AVA and AVA-Kinetics datasets, outperforming recent works which develop intricate models.
Adversarial patch attacks mislead neural networks by injecting adversarial pixels within a local region. Patch attacks can be highly effective in a variety of tasks and physically realizable via attachment (e.g. a sticker) to the real-world objects. Despite the diversity in attack patterns, adversarial patches tend to be highly textured and different in appearance from natural images. We exploit this property and present PatchZero, a general defense pipeline against white-box adversarial patches without retraining the downstream classifier or detector. Specifically, our defense detects adversaries at the pixel-level and "zeros out" the patch region by repainting with mean pixel values. We further design a two-stage adversarial training scheme to defend against the stronger adaptive attacks. PatchZero achieves SOTA defense performance on the image classification (ImageNet, RESISC45), object detection (PASCAL VOC), and video classification (UCF101) tasks with little degradation in benign performance. In addition, PatchZero transfers to different patch shapes and attack types.
Transferring knowledge from task-agnostic pre-trained deep models for downstream tasks is an important topic in computer vision research. Along with the growth of computational capacity, we now have open-source vision-language pre-trained models in large scales of the model architecture and amount of data. In this study, we focus on transferring knowledge for video classification tasks. Conventional methods randomly initialize the linear classifier head for vision classification, but they leave the usage of the text encoder for downstream visual recognition tasks undiscovered. In this paper, we revise the role of the linear classifier and replace the classifier with the different knowledge from pre-trained model. We utilize the well-pretrained language model to generate good semantic target for efficient transferring learning. The empirical study shows that our method improves both the performance and the training speed of video classification, with a negligible change in the model. Our simple yet effective tuning paradigm achieves state-of-the-art performance and efficient training on various video recognition scenarios, i.e., zero-shot, few-shot, general recognition. In particular, our paradigm achieves the state-of-the-art accuracy of 87.8% on Kinetics-400, and also surpasses previous methods by 20~50% absolute top-1 accuracy under zero-shot, few-shot settings on five popular video datasets. Code and models can be found at https://github.com/whwu95/Text4Vis .
Ultrasound examination is widely used in the clinical diagnosis of thyroid nodules (benign/malignant). However, the accuracy relies heavily on radiologist experience. Although deep learning techniques have been investigated for thyroid nodules recognition. Current solutions are mainly based on static ultrasound images, with limited temporal information used and inconsistent with clinical diagnosis. This paper proposes a novel method for the automated recognition of thyroid nodules through an exhaustive exploration of ultrasound videos and key-frames. We first propose a detection-localization framework to automatically identify the clinical key-frame with a typical nodule in each ultrasound video. Based on the localized key-frame, we develop a key-frame guided video classification model for thyroid nodule recognition. Besides, we introduce a motion attention module to help the network focus on significant frames in an ultrasound video, which is consistent with clinical diagnosis. The proposed thyroid nodule recognition framework is validated on clinically collected ultrasound videos, demonstrating superior performance compared with other state-of-the-art methods.
Recent efforts in interpretable deep learning models have shown that concept-based explanation methods achieve competitive accuracy with standard end-to-end models and enable reasoning and intervention about extracted high-level visual concepts from images, e.g., identifying the wing color and beak length for bird-species classification. However, these concept bottleneck models rely on a necessary and sufficient set of predefined concepts-which is intractable for complex tasks such as video classification. For complex tasks, the labels and the relationship between visual elements span many frames, e.g., identifying a bird flying or catching prey-necessitating concepts with various levels of abstraction. To this end, we present CoDEx, an automatic Concept Discovery and Extraction module that rigorously composes a necessary and sufficient set of concept abstractions for concept-based video classification. CoDEx identifies a rich set of complex concept abstractions from natural language explanations of videos-obviating the need to predefine the amorphous set of concepts. To demonstrate our method's viability, we construct two new public datasets that combine existing complex video classification datasets with short, crowd-sourced natural language explanations for their labels. Our method elicits inherent complex concept abstractions in natural language to generalize concept-bottleneck methods to complex tasks.
Adversarial training (AT) is a simple yet effective defense against adversarial attacks to image classification systems, which is based on augmenting the training set with attacks that maximize the loss. However, the effectiveness of AT as a defense for video classification has not been thoroughly studied. Our first contribution is to show that generating optimal attacks for video requires carefully tuning the attack parameters, especially the step size. Notably, we show that the optimal step size varies linearly with the attack budget. Our second contribution is to show that using a smaller (sub-optimal) attack budget at training time leads to a more robust performance at test time. Based on these findings, we propose three defenses against attacks with variable attack budgets. The first one, Adaptive AT, is a technique where the attack budget is drawn from a distribution that is adapted as training iterations proceed. The second, Curriculum AT, is a technique where the attack budget is increased as training iterations proceed. The third, Generative AT, further couples AT with a denoising generative adversarial network to boost robust performance. Experiments on the UCF101 dataset demonstrate that the proposed methods improve adversarial robustness against multiple attack types.
To build an artificial neural network like the biological intelligence system, recent works have unified numerous tasks into a generalist model, which can process various tasks with shared parameters and do not have any task-specific modules. While generalist models achieve promising results on various benchmarks, they have performance degradation on some tasks compared with task-specialized models. In this work, we find that interference among different tasks and modalities is the main factor to this phenomenon. To mitigate such interference, we introduce the Conditional Mixture-of-Experts (Conditional MoEs) to generalist models. Routing strategies under different levels of conditions are proposed to take both the training/inference cost and generalization ability into account. By incorporating the proposed Conditional MoEs, the recently proposed generalist model Uni-Perceiver can effectively mitigate the interference across tasks and modalities, and achieves state-of-the-art results on a series of downstream tasks via prompt tuning on 1% of downstream data. Moreover, the introduction of Conditional MoEs still holds the generalization ability of generalist models to conduct zero-shot inference on new tasks, e.g., video-text retrieval and video caption. Code and pre-trained generalist models shall be released.
Most existing state-of-the-art video classification methods assume that the training data obey a uniform distribution. However, video data in the real world typically exhibit an imbalanced long-tailed class distribution, resulting in a model bias towards head class and relatively low performance on tail class. While the current long-tailed classification methods usually focus on image classification, adapting it to video data is not a trivial extension. We propose an end-to-end multi-expert distribution calibration method to address these challenges based on two-level distribution information. The method jointly considers the distribution of samples in each class (intra-class distribution) and the overall distribution of diverse data (inter-class distribution) to solve the issue of imbalanced data under long-tailed distribution. By modeling the two-level distribution information, the model can jointly consider the head classes and the tail classes and significantly transfer the knowledge from the head classes to improve the performance of the tail classes. Extensive experiments verify that our method achieves state-of-the-art performance on the long-tailed video classification task.
Weakly-supervised temporal action localization (WTAL) intends to detect action instances with only weak supervision, e.g., video-level labels. The current~\textit{de facto} pipeline locates action instances by thresholding and grouping continuous high-score regions on temporal class activation sequences. In this route, the capacity of the model to recognize the relationships between adjacent snippets is of vital importance which determines the quality of the action boundaries. However, it is error-prone since the variations between adjacent snippets are typically subtle, and unfortunately this is overlooked in the literature. To tackle the issue, we propose a novel WTAL approach named Convex Combination Consistency between Neighbors (C$^3$BN). C$^3$BN consists of two key ingredients: a micro data augmentation strategy that increases the diversity in-between adjacent snippets by convex combination of adjacent snippets, and a macro-micro consistency regularization that enforces the model to be invariant to the transformations~\textit{w.r.t.} video semantics, snippet predictions, and snippet representations. Consequently, fine-grained patterns in-between adjacent snippets are enforced to be explored, thereby resulting in a more robust action boundary localization. Experimental results demonstrate the effectiveness of C$^3$BN on top of various baselines for WTAL with video-level and point-level supervisions. Code is at https://github.com/Qinying-Liu/C3BN.
Attention mechanisms have significantly boosted the performance of video classification neural networks thanks to the utilization of perspective contexts. However, the current research on video attention generally focuses on adopting a specific aspect of contexts (e.g., channel, spatial/temporal, or global context) to refine the features and neglects their underlying correlation when computing attentions. This leads to incomplete context utilization and hence bears the weakness of limited performance improvement. To tackle the problem, this paper proposes an efficient attention-in-attention (AIA) method for element-wise feature refinement, which investigates the feasibility of inserting the channel context into the spatio-temporal attention learning module, referred to as CinST, and also its reverse variant, referred to as STinC. Specifically, we instantiate the video feature contexts as dynamics aggregated along a specific axis with global average and max pooling operations. The workflow of an AIA module is that the first attention block uses one kind of context information to guide the gating weights calculation of the second attention that targets at the other context. Moreover, all the computational operations in attention units act on the pooled dimension, which results in quite few computational cost increase ($<$0.02\%). To verify our method, we densely integrate it into two classical video network backbones and conduct extensive experiments on several standard video classification benchmarks. The source code of our AIA is available at \url{https://github.com/haoyanbin918/Attention-in-Attention}.
Assuming the source label space subsumes the target one, Partial Video Domain Adaptation (PVDA) is a more general and practical scenario for cross-domain video classification problems. The key challenge of PVDA is to mitigate the negative transfer caused by the source-only outlier classes. To tackle this challenge, a crucial step is to aggregate target predictions to assign class weights by up-weighing target classes and down-weighing outlier classes. However, the incorrect predictions of class weights can mislead the network and lead to negative transfer. Previous works improve the class weight accuracy by utilizing temporal features and attention mechanisms, but these methods may fall short when trying to generate accurate class weight when domain shifts are significant, as in most real-world scenarios. To deal with these challenges, we propose the Multi-modality Cluster-calibrated partial Adversarial Network (MCAN). MCAN enhances video feature extraction with multi-modal features from multiple temporal scales to form more robust overall features. It utilizes a novel class weight calibration method to alleviate the negative transfer caused by incorrect class weights. The calibration method tries to identify and weigh correct and incorrect predictions using distributional information implied by unsupervised clustering. Extensive experiments are conducted on prevailing PVDA benchmarks, and the proposed MCAN achieves significant improvements when compared to state-of-the-art PVDA methods.
Learning an egocentric action recognition model from video data is challenging due to distractors (e.g., irrelevant objects) in the background. Further integrating object information into an action model is hence beneficial. Existing methods often leverage a generic object detector to identify and represent the objects in the scene. However, several important issues remain. Object class annotations of good quality for the target domain (dataset) are still required for learning good object representation. Besides, previous methods deeply couple the existing action models and need to retrain them jointly with object representation, leading to costly and inflexible integration. To overcome both limitations, we introduce Self-Supervised Learning Over Sets (SOS), an approach to pre-train a generic Objects In Contact (OIC) representation model from video object regions detected by an off-the-shelf hand-object contact detector. Instead of augmenting object regions individually as in conventional self-supervised learning, we view the action process as a means of natural data transformations with unique spatio-temporal continuity and exploit the inherent relationships among per-video object sets. Extensive experiments on two datasets, EPIC-KITCHENS-100 and EGTEA, show that our OIC significantly boosts the performance of multiple state-of-the-art video classification models.
Most modern video recognition models are designed to operate on short video clips (e.g., 5-10s in length). Thus, it is challenging to apply such models to long movie understanding tasks, which typically require sophisticated long-range temporal reasoning. The recently introduced video transformers partially address this issue by using long-range temporal self-attention. However, due to the quadratic cost of self-attention, such models are often costly and impractical to use. Instead, we propose ViS4mer, an efficient long-range video model that combines the strengths of self-attention and the recently introduced structured state-space sequence (S4) layer. Our model uses a standard Transformer encoder for short-range spatiotemporal feature extraction, and a multi-scale temporal S4 decoder for subsequent long-range temporal reasoning. By progressively reducing the spatiotemporal feature resolution and channel dimension at each decoder layer, ViS4mer learns complex long-range spatiotemporal dependencies in a video. Furthermore, ViS4mer is $2.63\times$ faster and requires $8\times$ less GPU memory than the corresponding pure self-attention-based model. Additionally, ViS4mer achieves state-of-the-art results in $6$ out of $9$ long-form movie video classification tasks on the Long Video Understanding (LVU) benchmark. Furthermore, we show that our approach successfully generalizes to other domains, achieving competitive results on the Breakfast and the COIN procedural activity datasets. The code is publicly available at: https://github.com/md-mohaiminul/ViS4mer.
Video classification systems are vulnerable to adversarial attacks, which can create severe security problems in video verification. Current black-box attacks need a large number of queries to succeed, resulting in high computational overhead in the process of attack. On the other hand, attacks with restricted perturbations are ineffective against defenses such as denoising or adversarial training. In this paper, we focus on unrestricted perturbations and propose StyleFool, a black-box video adversarial attack via style transfer to fool the video classification system. StyleFool first utilizes color theme proximity to select the best style image, which helps avoid unnatural details in the stylized videos. Meanwhile, the target class confidence is additionally considered in targeted attacks to influence the output distribution of the classifier by moving the stylized video closer to or even across the decision boundary. A gradient-free method is then employed to further optimize the adversarial perturbations. We carry out extensive experiments to evaluate StyleFool on two standard datasets, UCF-101 and HMDB-51. The experimental results demonstrate that StyleFool outperforms the state-of-the-art adversarial attacks in terms of both the number of queries and the robustness against existing defenses. Moreover, 50% of the stylized videos in untargeted attacks do not need any query since they can already fool the video classification model. Furthermore, we evaluate the indistinguishability through a user study to show that the adversarial samples of StyleFool look imperceptible to human eyes, despite unrestricted perturbations.