Human-Object Interaction (HOI) detection aims to learn how human interacts with surrounding objects. Previous HOI detection frameworks simultaneously detect human, objects and their corresponding interactions by using a predictor. Using only one shared predictor cannot differentiate the attentive field of instance-level prediction and relation-level prediction. To solve this problem, we propose a new transformer-based method named Parallel Reasoning Network(PR-Net), which constructs two independent predictors for instance-level localization and relation-level understanding. The former predictor concentrates on instance-level localization by perceiving instances' extremity regions. The latter broadens the scope of relation region to reach a better relation-level semantic understanding. Extensive experiments and analysis on HICO-DET benchmark exhibit that our PR-Net effectively alleviated this problem. Our PR-Net has achieved competitive results on HICO-DET and V-COCO benchmarks.
Recent attempts mainly focus on learning deep representations for each video individually under the episodic meta-learning regime and then performing temporal alignment to match query and support videos. However, they still suffer from two drawbacks: (i) learning individual features without considering the entire task may result in limited representation capability, and (ii) existing alignment strategies are sensitive to noises and misaligned instances. To handle the two limitations, we propose a novel Hybrid Relation guided temporal Set Matching (HyRSM++) approach for few-shot action recognition. The core idea of HyRSM++ is to integrate all videos within the task to learn discriminative representations and involve a robust matching technique. To be specific, HyRSM++ consists of two key components, a hybrid relation module and a temporal set matching metric. Given the basic representations from the feature extractor, the hybrid relation module is introduced to fully exploit associated relations within and cross videos in an episodic task and thus can learn task-specific embeddings. Subsequently, in the temporal set matching metric, we carry out the distance measure between query and support videos from a set matching perspective and design a Bi-MHM to improve the resilience to misaligned instances. In addition, we explicitly exploit the temporal coherence in videos to regularize the matching process. Furthermore, we extend the proposed HyRSM++ to deal with the more challenging semi-supervised few-shot action recognition and unsupervised few-shot action recognition tasks. Experimental results on multiple benchmarks demonstrate that our method achieves state-of-the-art performance under various few-shot settings. The source code is available at https://github.com/alibaba-mmai-research/HyRSMPlusPlus.
Recent incremental learning for action recognition usually stores representative videos to mitigate catastrophic forgetting. However, only a few bulky videos can be stored due to the limited memory. To address this problem, we propose FrameMaker, a memory-efficient video class-incremental learning approach that learns to produce a condensed frame for each selected video. Specifically, FrameMaker is mainly composed of two crucial components: Frame Condensing and Instance-Specific Prompt. The former is to reduce the memory cost by preserving only one condensed frame instead of the whole video, while the latter aims to compensate the lost spatio-temporal details in the Frame Condensing stage. By this means, FrameMaker enables a remarkable reduction in memory but keep enough information that can be applied to following incremental tasks. Experimental results on multiple challenging benchmarks, i.e., HMDB51, UCF101 and Something-Something V2, demonstrate that FrameMaker can achieve better performance to recent advanced methods while consuming only 20% memory. Additionally, under the same memory consumption conditions, FrameMaker significantly outperforms existing state-of-the-arts by a convincing margin.
Standard approaches for video recognition usually operate on the full input videos, which is inefficient due to the widely present spatio-temporal redundancy in videos. Recent progress in masked video modelling, i.e., VideoMAE, has shown the ability of vanilla Vision Transformers (ViT) to complement spatio-temporal contexts given only limited visual contents. Inspired by this, we propose propose Masked Action Recognition (MAR), which reduces the redundant computation by discarding a proportion of patches and operating only on a part of the videos. MAR contains the following two indispensable components: cell running masking and bridging classifier. Specifically, to enable the ViT to perceive the details beyond the visible patches easily, cell running masking is presented to preserve the spatio-temporal correlations in videos, which ensures the patches at the same spatial location can be observed in turn for easy reconstructions. Additionally, we notice that, although the partially observed features can reconstruct semantically explicit invisible patches, they fail to achieve accurate classification. To address this, a bridging classifier is proposed to bridge the semantic gap between the ViT encoded features for reconstruction and the features specialized for classification. Our proposed MAR reduces the computational cost of ViT by 53% and extensive experiments show that MAR consistently outperforms existing ViT models with a notable margin. Especially, we found a ViT-Large trained by MAR outperforms the ViT-Huge trained by a standard training scheme by convincing margins on both Kinetics-400 and Something-Something v2 datasets, while our computation overhead of ViT-Large is only 14.5% of ViT-Huge.
This technical report presents our first place winning solution for temporal action detection task in CVPR-2022 AcitivityNet Challenge. The task aims to localize temporal boundaries of action instances with specific classes in long untrimmed videos. Recent mainstream attempts are based on dense boundary matchings and enumerate all possible combinations to produce proposals. We argue that the generated proposals contain rich contextual information, which may benefits detection confidence prediction. To this end, our method mainly consists of the following three steps: 1) action classification and feature extraction by Slowfast, CSN, TimeSformer, TSP, I3D-flow, VGGish-audio, TPN and ViViT; 2) proposal generation. Our proposed Context-aware Proposal Network (CPN) builds on top of BMN, GTAD and PRN to aggregate contextual information by randomly masking some proposal features. 3) action detection. The final detection prediction is calculated by assigning the proposals with corresponding video-level classifcation results. Finally, we ensemble the results under different feature combination settings and achieve 45.8% performance on the test set, which improves the champion result in CVPR-2021 ActivityNet Challenge by 1.1% in terms of average mAP.
Current few-shot action recognition methods reach impressive performance by learning discriminative features for each video via episodic training and designing various temporal alignment strategies. Nevertheless, they are limited in that (a) learning individual features without considering the entire task may lose the most relevant information in the current episode, and (b) these alignment strategies may fail in misaligned instances. To overcome the two limitations, we propose a novel Hybrid Relation guided Set Matching (HyRSM) approach that incorporates two key components: hybrid relation module and set matching metric. The purpose of the hybrid relation module is to learn task-specific embeddings by fully exploiting associated relations within and cross videos in an episode. Built upon the task-specific features, we reformulate distance measure between query and support videos as a set matching problem and further design a bidirectional Mean Hausdorff Metric to improve the resilience to misaligned instances. By this means, the proposed HyRSM can be highly informative and flexible to predict query categories under the few-shot settings. We evaluate HyRSM on six challenging benchmarks, and the experimental results show its superiority over the state-of-the-art methods by a convincing margin. Project page: https://hyrsm-cvpr2022.github.io/.
Natural videos provide rich visual contents for self-supervised learning. Yet most existing approaches for learning spatio-temporal representations rely on manually trimmed videos, leading to limited diversity in visual patterns and limited performance gain. In this work, we aim to learn representations by leveraging more abundant information in untrimmed videos. To this end, we propose to learn a hierarchy of consistencies in videos, i.e., visual consistency and topical consistency, corresponding respectively to clip pairs that tend to be visually similar when separated by a short time span and share similar topics when separated by a long time span. Specifically, a hierarchical consistency learning framework HiCo is presented, where the visually consistent pairs are encouraged to have the same representation through contrastive learning, while the topically consistent pairs are coupled through a topical classifier that distinguishes whether they are topic related. Further, we impose a gradual sampling algorithm for proposed hierarchical consistency learning, and demonstrate its theoretical superiority. Empirically, we show that not only HiCo can generate stronger representations on untrimmed videos, it also improves the representation quality when applied to trimmed videos. This is in contrast to standard contrastive learning that fails to learn appropriate representations from untrimmed videos.
Currently, for crowd counting, the fully supervised methods via density map estimation are the mainstream research directions. However, such methods need location-level annotation of persons in an image, which is time-consuming and laborious. Therefore, the weakly supervised method just relying upon the count-level annotation is urgently needed. Since CNN is not suitable for modeling the global context and the interactions between image patches, crowd counting with weakly supervised learning via CNN generally can not show good performance. The weakly supervised model via Transformer was sequentially proposed to model the global context and learn contrast features. However, the transformer directly partitions the crowd images into a series of tokens, which may not be a good choice due to each pedestrian being an independent individual, and the parameter number of the network is very large. Hence, we propose a Joint CNN and Transformer Network (JCTNet) via weakly supervised learning for crowd counting in this paper. JCTNet consists of three parts: CNN feature extraction module (CFM), Transformer feature extraction module (TFM), and counting regression module (CRM). In particular, the CFM extracts crowd semantic information features, then sends their patch partitions to TRM for modeling global context, and CRM is used to predict the number of people. Extensive experiments and visualizations demonstrate that JCTNet can effectively focus on the crowd regions and obtain superior weakly supervised counting performance on five mainstream datasets. The number of parameters of the model can be reduced by about 67%~73% compared with the pure Transformer works. We also tried to explain the phenomenon that a model constrained only by count-level annotations can still focus on the crowd regions. We believe our work can promote further research in this field.
Recently, many approaches tackle the Unsupervised Domain Adaptive person re-identification (UDA re-ID) problem through pseudo-label-based contrastive learning. During training, a uni-centroid representation is obtained by simply averaging all the instance features from a cluster with the same pseudo label. However, a cluster may contain images with different identities (label noises) due to the imperfect clustering results, which makes the uni-centroid representation inappropriate. In this paper, we present a novel Multi-Centroid Memory (MCM) to adaptively capture different identity information within the cluster. MCM can effectively alleviate the issue of label noises by selecting proper positive/negative centroids for the query image. Moreover, we further propose two strategies to improve the contrastive learning process. First, we present a Domain-Specific Contrastive Learning (DSCL) mechanism to fully explore intradomain information by comparing samples only from the same domain. Second, we propose Second-Order Nearest Interpolation (SONI) to obtain abundant and informative negative samples. We integrate MCM, DSCL, and SONI into a unified framework named Multi-Centroid Representation Network (MCRN). Extensive experiments demonstrate the superiority of MCRN over state-of-the-art approaches on multiple UDA re-ID tasks and fully unsupervised re-ID tasks.
This paper tackles the Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR) problem from the viewpoint of cross-modality metric learning. This task has two characteristics: 1) the zero-shot setting requires a metric space with good within-class compactness and the between-class discrepancy for recognizing the novel classes and 2) the sketch query and the photo gallery are in different modalities. The metric learning viewpoint benefits ZS-SBIR from two aspects. First, it facilitates improvement through recent good practices in deep metric learning (DML). By combining two fundamental learning approaches in DML, e.g., classification training and pairwise training, we set up a strong baseline for ZS-SBIR. Without bells and whistles, this baseline achieves competitive retrieval accuracy. Second, it provides an insight that properly suppressing the modality gap is critical. To this end, we design a novel method named Modality-Aware Triplet Hard Mining (MATHM). MATHM enhances the baseline with three types of pairwise learning, e.g., a cross-modality sample pair, a within-modality sample pair, and their combination.\We also design an adaptive weighting method to balance these three components during training dynamically. Experimental results confirm that MATHM brings another round of significant improvement based on the strong baseline and sets up new state-of-the-art performance. For example, on the TU-Berlin dataset, we achieve 47.88+2.94% mAP@all and 58.28+2.34% Prec@100. Code will be publicly available at: https://github.com/huangzongheng/MATHM.