Students’ emotional engagement in the classroom can be evaluated through their facial emotions. However, facial emotion recognition remains challenging due to the uncertainty problem, which is routinely caused by factors such as low resolution, intra-class variability, and inter-class similarity. To deal with this problem, we propose a class activation regularization network, which includes two main modules: class activation dropping module (CADM) and hard sample mining module (HSMM). Our key idea is to perform consistency learning between the original images and class-uncertain regions, thereby enabling models to perceive uncertainty of emotional semantics. In CADM, the class-specific attention is obtained using class activation mapping. Then, the attention is partly discarded by imitating dropout mechanism to obtain class-uncertain regions. In HSMM, the hard samples with high uncertainties are identified based on the class-uncertain regions and confidence thresholds. For the hard samples, regularization loss is applied to align the probability distributions between the original images and the class-uncertain regions, thereby improving the network’s awareness of emotional semantics and its ability to handle uncertainty. Extensive experiments on three datasets demonstrate the competitiveness of the proposed method. In addition, our method can be deployed in the classroom for students’ emotional engagement assessment.

Abstract Facial emotion recognition in the wild is an important task in computer vision, but it still remains challenging since the influence of backgrounds, occlusions and illumination variations in facial images, as well as the ambiguity of expressions. This paper proposes a harmonious mutual learning framework for emotion recognition, mainly through utilizing attention mechanisms and probability distributions without utilizing additional information. Specifically, this paper builds an architecture with two emotion recognition networks and makes progressive cooperation and interaction between them. We first integrate self-mutual attention module into the backbone to learn discriminative features against the influence from emotion-irrelevant facial information. In this process, we deploy spatial attention module and convolutional block attention module for the two networks respectively, guiding to enhanced and supplementary learning of attention. Further, in the classification head, we propose to learn the latent ground-truth emotion probability distributions using softmax function with temperature to characterize the expression ambiguity. On this basis, a probability distribution distillation learning module is constructed to perform class semantic interaction using bi-directional KL loss, allowing mutual calibration for the two networks. Experimental results on three public datasets show the superiority of the proposed method compared to state-of-the-art ones.

Applying large scale pre-trained image-language model to video-language tasks has recently become a trend, which brings two challenges. One is how to effectively transfer knowledge from static images to dynamic videos, and the other is how to deal with the prohibitive cost of fully fine-tuning due to growing model size. Existing works that attempt to realize parameter-efficient image-language to video-language transfer learning can be categorized into two types: 1) appending a sequence of temporal transformer blocks after the 2D Vision Transformer (ViT), and 2) inserting a temporal block into the ViT architecture. While these two types of methods only require fine-tuning the newly added components, there are still many parameters to update, and they are only validated on a single video-language task. In this work, based on our analysis of the core ideas of different temporal modeling components in existing approaches, we propose a token mixing strategy to enable cross-frame interactions, which enables transferring from the pre-trained image-language model to video-language tasks through selecting and mixing a key set and a value set from the input video samples. As token mixing does not require the addition of any components or modules, we can directly partially fine-tune the pre-trained image-language model to achieve parameter-efficiency. We carry out extensive experiments to compare our proposed token mixing method with other parameter-efficient transfer learning methods. Our token mixing method outperforms other methods on both understanding tasks and generation tasks. Besides, our method achieves new records on multiple video-language tasks. The code is available at https://github.com/yuqi657/video_language_model.

We present a novel network to transfer the image-language pre-trained model to video-text retrieval in an end-to-end manner. Leading approaches in the domain of video-and-language learning try to distill the spatio-temporal video features and multi-modal interaction between videos and language from a large-scale video-text dataset. Differently, we leverage the pre-trained image-language model, and simplify it as a two-stage framework including co-learning of image and text, and enhancing temporal relations between video frames and video-text respectively. Specifically, based on the spatial semantics captured by Contrastive Language-Image Pre-training (CLIP) model, our model involves a Temporal Difference Block (TDB) to capture motions at fine temporal video frames, and a Temporal Alignment Block (TAB) to re-align the tokens of video clips and phrases and enhance the cross-modal correlation. These two temporal blocks efficiently realize video-language learning and enable the proposed model to scale well on comparatively small datasets. We conduct extensive experimental studies including ablation studies and comparisons with existing SOTA methods, and our proposed approach outperforms them on the popularly-employed text-to-video and video-to-text retrieval benchmarks, including MSR-VTT, MSVD, LSMDC, and VATEX.