Semi-supervised video learning is an increasingly popular approach for improving video understanding tasks by utilizing large-scale unlabeled videos along with a few labels. Recent studies have shown that multimodal contrastive learning and consistency regularization are effective techniques for generating high-quality pseudo-labels for semi-supervised action recognition. However, existing pseudo-labeling approaches are solely based on the model's class predictions and can suffer from confirmation biases due to the accumulation of false predictions. To address this issue, we propose exploiting audio-visual feature correlations to achieve high-quality pseudo-labels instead of relying on model confidence. To achieve this goal, we introduce Audio-visual Contrastive and Consistency Learning (AvCLR) for semi-supervised action recognition. AvCLR generates reliable pseudo-labels from audio-visual feature correlations using deep embedded clustering to mitigate confirmation biases. Additionally, AvCLR introduces two contrastive modules: intra-modal contrastive learning (ImCL) and cross-modal contrastive learning (XmCL) to discover complementary information from audio-visual alignments. The ImCL module learns informative representations within audio and video independently, while the XmCL module aims to leverage global high-level features of audio-visual information. Furthermore, the XmCL is constrained by introducing intra-instance negatives from one modality to the other. We jointly optimize the model with ImCL, XmCL, and consistency regularization in an end-to-end semi-supervised manner. Experimental results have demonstrated that the proposed AvCLR framework is effective in reducing confirmation biases and outperforms existing confidence-based semi-supervised action recognition methods.

Self-supervised contrastive learning has shown a significant improvement in performance for action recognition tasks by discovering useful signals from unlabeled videos. Nevertheless, the unique features of existing video benchmark datasets have led the learned video representations to be contextually biased toward dominant backgrounds and scene correlations. Thus, ultimately leading to poor generalizations on scene-invariant action recognition. Therefore, we propose Actor-aware Self-supervised Learning for Semi-supervised Video Representation Learning (ActorSL). We aligned localized actors and their corresponding scene information to encourage the model to learn discriminative regions and mitigate the model’s dependency on the video background during contrastive training. Furthermore, we present an inter-video Background Mixing (iBM) augmentation strategy to introduce scene consistency into the model. We patch inter-video crops of four randomly selected frames for iBM to create a unique frame for each video. The patched frame is blended with the target video frames to generate a spatially augmented sample. Then, the actor-scene aligned features and features of iBM-augmented videos are utilized to optimize contrastive loss and consistency regularization jointly in a semi-supervised way. Moreover, iBM combines the one-hot-encoded labels of patches with the label of the target video as a label smoothing regularizer to soften the decision boundaries of the semi-supervised model. Our experimental results reveal that, ActorSL notably improved current state-of-the-art semi-supervised methods on the Kinetics-400, UCF101, and HMDB51 datasets under a low-label regime. Code released athttps://github.com/Endarzboy/ActorSL.

Self-supervised learning has considerably improved video representation learning by discovering supervisory signals automatically from unlabeled videos. However, due to the scene-biased nature of existing video datasets, the current methods are biased to the dominant scene context during action inference. Hence, this paper proposes Background Patching (BP), a scene-debiasing augmentation strategy to alleviate the model reliance on the video background in a self-supervised contrastive manner. The BP reduces the negative influence of the video background by mixing a randomly patched frame to the video background. BP randomly crops four frames from four different videos and patches them to construct a new frame for each video separately. The patched frame is mixed with all frames of the target video to produce a spatially distorted video sample. Then, we use existing self-supervised contrastive frameworks to pull representations of the distorted and original videos closer together. Moreover, BP mixes the semantic labels of patches with the target video's label, resulting in the regularization of the contrastive model to soften the decision boundaries in the embedding space. Therefore, the model is explicitly constrained to suppress the background influence by emphasizing more on the motion changes. The extensive experimental results show that our BP significantly improved the performance of various video understanding downstream tasks including action recognition, action detection, and video retrieval.

The unique features of existing video datasets have led self-supervised contrastive learning to scene correlations and background biases, resulting in poor generalization in scene-invariant action recognition. Therefore, we propose Actor-aware Contrastive Learning for semi-supervised action recognition (ActorCLR). We employ localized actors to encourage the model to learn discriminative regions and mitigate the model's reliance on the video background during contrastive training. Furthermore, we introduce Inter-video Background Mixing (iBM) augmentation strategy to inject scene-invariance into the model. For iBM, we patch inter-video crops of four randomly selected frames to create a distinct frame for each video individually. The patched frame is mixed with the target video frames to produce a spatially distorted sample. Then, we jointly optimize contrastive loss and consistency regularization with localized actors and corresponding iBM-augmented videos in a semi-supervised manner. iBM also mixes the one-hot-encoded labels of patches with the target video's label, which softens the decision boundaries of the semi-supervised model. Our experimental results show that ActorCLR significantly improved action recognition on Kinetics-400, UCF101, and HMDB51 datasets under a low-label regime.