Contrastive Language-Image Pretraining (CLIP) has shown remarkable open-vocabulary abilities across various image understanding tasks. Building upon this impressive success, recent pioneer works have proposed to adapt the powerful CLIP to video data, leading to efficient and effective video learners for open-vocabulary action recognition. Inspired by the fact that humans perform actions in diverse environments, our work delves into an intriguing question: Can CLIP-based video learners effectively generalize to video domains they have not encountered during training? To answer this, we establish a CROSS-domain Open-Vocabulary Action recognition benchmark named XOV-Action, and conduct a comprehensive evaluation of five state-of-the-art CLIP-based video learners under various types of domain gaps. Our evaluation demonstrates that previous methods exhibit limited action recognition performance in unseen video domains, revealing potential challenges of the cross-domain open-vocabulary action recognition task. To address this task, our work focuses on a critical challenge, namely scene bias, and we accordingly contribute a novel scene-aware video-text alignment method. Our key idea is to distinguish video representations apart from scene-encoded text representations, aiming to learn scene-agnostic video representations for recognizing actions across domains. Extensive experimental results demonstrate the effectiveness of our method. The benchmark and code will be available at https://github.com/KunyuLin/XOV-Action/.
Zero-shot action recognition (ZSAR) aims to learn an alignment model between videos and class descriptions of seen actions that is transferable to unseen actions. The text queries (class descriptions) used in existing ZSAR works, however, are often short action names that fail to capture the rich semantics in the videos, leading to misalignment. With the intuition that video content descriptions (e.g., video captions) can provide rich contextual information of visual concepts in videos, we propose to utilize human annotated video descriptions to enrich the semantics of the class descriptions of each action. However, all existing action video description datasets are limited in terms of the number of actions, the semantics of video descriptions, etc. To this end, we collect a large-scale action video descriptions dataset named ActionHub, which covers a total of 1,211 common actions and provides 3.6 million action video descriptions. With the proposed ActionHub dataset, we further propose a novel Cross-modality and Cross-action Modeling (CoCo) framework for ZSAR, which consists of a Dual Cross-modality Alignment module and a Cross-action Invariance Mining module. Specifically, the Dual Cross-modality Alignment module utilizes both action labels and video descriptions from ActionHub to obtain rich class semantic features for feature alignment. The Cross-action Invariance Mining module exploits a cycle-reconstruction process between the class semantic feature spaces of seen actions and unseen actions, aiming to guide the model to learn cross-action invariant representations. Extensive experimental results demonstrate that our CoCo framework significantly outperforms the state-of-the-art on three popular ZSAR benchmarks (i.e., Kinetics-ZSAR, UCF101 and HMDB51) under two different learning protocols in ZSAR. We will release our code, models, and the proposed ActionHub dataset.
Developing end-to-end models for long-video action understanding tasks presents significant computational and memory challenges. Existing works generally build models on long-video features extracted by off-the-shelf action recognition models, which are trained on short-video datasets in different domains, making the extracted features suffer domain discrepancy. To avoid this, action recognition models can be end-to-end trained on clips, which are trimmed from long videos and labeled using action interval annotations. Such fully supervised annotations are expensive to collect. Thus, a weakly supervised method is needed for long-video action understanding at scale. Under the weak supervision setting, action labels are provided for the whole video without precise start and end times of the action clip. To this end, we propose an AdaFocus framework. AdaFocus estimates the spike-actionness and temporal positions of actions, enabling it to adaptively focus on action clips that facilitate better training without the need for precise annotations. Experiments on three long-video datasets show its effectiveness. Remarkably, on two of datasets, models trained with AdaFocus under weak supervision outperform those trained under full supervision. Furthermore, we form a weakly supervised feature extraction pipeline with our AdaFocus, which enables significant improvements on three long-video action understanding tasks.
Video domain generalization aims to learn generalizable video classification models for unseen target domains by training in a source domain. A critical challenge of video domain generalization is to defend against the heavy reliance on domain-specific cues extracted from the source domain when recognizing target videos. To this end, we propose to perceive diverse spatial-temporal cues in videos, aiming to discover potential domain-invariant cues in addition to domain-specific cues. We contribute a novel model named Spatial-Temporal Diversification Network (STDN), which improves the diversity from both space and time dimensions of video data. First, our STDN proposes to discover various types of spatial cues within individual frames by spatial grouping. Then, our STDN proposes to explicitly model spatial-temporal dependencies between video contents at multiple space-time scales by spatial-temporal relation modeling. Extensive experiments on three benchmarks of different types demonstrate the effectiveness and versatility of our approach.
In this work, we focus on the task of procedure planning from instructional videos with text supervision, where a model aims to predict an action sequence to transform the initial visual state into the goal visual state. A critical challenge of this task is the large semantic gap between observed visual states and unobserved intermediate actions, which is ignored by previous works. Specifically, this semantic gap refers to that the contents in the observed visual states are semantically different from the elements of some action text labels in a procedure. To bridge this semantic gap, we propose a novel event-guided paradigm, which first infers events from the observed states and then plans out actions based on both the states and predicted events. Our inspiration comes from that planning a procedure from an instructional video is to complete a specific event and a specific event usually involves specific actions. Based on the proposed paradigm, we contribute an Event-guided Prompting-based Procedure Planning (E3P) model, which encodes event information into the sequential modeling process to support procedure planning. To further consider the strong action associations within each event, our E3P adopts a mask-and-predict approach for relation mining, incorporating a probabilistic masking scheme for regularization. Extensive experiments on three datasets demonstrate the effectiveness of our proposed model.
As a de facto solution, the vanilla Vision Transformers (ViTs) are encouraged to model long-range dependencies between arbitrary image patches while the global attended receptive field leads to quadratic computational cost. Another branch of Vision Transformers exploits local attention inspired by CNNs, which only models the interactions between patches in small neighborhoods. Although such a solution reduces the computational cost, it naturally suffers from small attended receptive fields, which may limit the performance. In this work, we explore effective Vision Transformers to pursue a preferable trade-off between the computational complexity and size of the attended receptive field. By analyzing the patch interaction of global attention in ViTs, we observe two key properties in the shallow layers, namely locality and sparsity, indicating the redundancy of global dependency modeling in shallow layers of ViTs. Accordingly, we propose Multi-Scale Dilated Attention (MSDA) to model local and sparse patch interaction within the sliding window. With a pyramid architecture, we construct a Multi-Scale Dilated Transformer (DilateFormer) by stacking MSDA blocks at low-level stages and global multi-head self-attention blocks at high-level stages. Our experiment results show that our DilateFormer achieves state-of-the-art performance on various vision tasks. On ImageNet-1K classification task, DilateFormer achieves comparable performance with 70% fewer FLOPs compared with existing state-of-the-art models. Our DilateFormer-Base achieves 85.6% top-1 accuracy on ImageNet-1K classification task, 53.5% box mAP/46.1% mask mAP on COCO object detection/instance segmentation task and 51.1% MS mIoU on ADE20K semantic segmentation task.
Deep neural networks only learn to map in-distribution inputs to their corresponding ground truth labels in the training phase without differentiating out-of-distribution samples from in-distribution ones. This results from the assumption that all samples are independent and identically distributed without distributional distinction. Therefore, a pretrained network learned from the in-distribution samples treats out-of-distribution samples as in-distribution and makes high-confidence predictions on them in the test phase. To address this issue, we draw out-of-distribution samples from the vicinity distribution of training in-distribution samples for learning to reject the prediction on out-of-distribution inputs. A \textit{Cross-class Vicinity Distribution} is introduced by assuming that an out-of-distribution sample generated by mixing multiple in-distribution samples does not share the same classes of its constituents. We thus improve the discriminability of a pretrained network by finetuning it with out-of-distribution samples drawn from the cross-class vicinity distribution, where each out-of-distribution input corresponds to a complementary label. Experiments on various in-/out-of-distribution datasets show that the proposed method significantly outperforms existing methods in improving the capacity of discriminating between in- and out-of-distribution samples.
When there is a discrepancy between in-distribution (ID) samples and out-of-distribution (OOD) samples, deep neural networks trained on ID samples suffer from high-confidence prediction on OOD samples. This is primarily caused by unavailable OOD samples to constrain the networks in the training process. To improve the OOD sensitivity of deep networks, several state-of-the-art methods introduce samples from other real-world datasets as OOD samples to the training process and assign manually-determined labels to these OOD samples. However, they sacrifice the classification accuracy because the unreliable labeling of OOD samples would disrupt ID classification. To balance ID generalization and OOD detection, a major challenge to tackle is to make OOD samples compatible with ID ones, which is addressed by our proposed \textit{supervision adaptation} method in this paper to define adaptive supervision information for OOD samples. First, by measuring the dependency between ID samples and their labels through mutual information, we reveal the form of the supervision information in terms of the negative probabilities of all classes. Second, after exploring the data correlations between ID and OOD samples by solving multiple binary regression problems, we estimate the supervision information to make ID classes more separable. We perform experiments on four advanced network architectures with two ID datasets and eleven OOD datasets to demonstrate the balancing effect of our supervision adaptation method in achieving both the ID classification ability and the OOD detection capacity.
An explicit discriminator trained on observable in-distribution (ID) samples can make high-confidence prediction on out-of-distribution (OOD) samples due to its distributional vulnerability. This is primarily caused by the limited ID samples observable for training discriminators when OOD samples are unavailable. To address this issue, the state-of-the-art methods train the discriminator with OOD samples generated by general assumptions without considering the data and network characteristics. However, different network architectures and training ID datasets may cause diverse vulnerabilities, and the generated OOD samples thus usually misaddress the specific distributional vulnerability of the explicit discriminator. To reveal and patch the distributional vulnerabilities, we propose a novel method of \textit{fine-tuning explicit discriminators by implicit generators} (FIG). According to the Shannon entropy, an explicit discriminator can construct its corresponding implicit generator to generate specific OOD samples without extra training costs. A Langevin Dynamic sampler then draws high-quality OOD samples from the generator to reveal the vulnerability. Finally, a regularizer, constructed according to the design principle of the implicit generator, patches the distributional vulnerability by encouraging those generated OOD samples with high entropy. Our experiments on four networks, four ID datasets and seven OOD datasets demonstrate that FIG achieves state-of-the-art OOD detection performance and maintains a competitive classification capability.