Vision-Language Models (VLMs) bring powerful understanding and reasoning capabilities to multimodal tasks. Meanwhile, the great need for capable aritificial intelligence on mobile devices also arises, such as the AI assistant software. Some efforts try to migrate VLMs to edge devices to expand their application scope. Simplifying the model structure is a common method, but as the model shrinks, the trade-off between performance and size becomes more and more difficult. Knowledge distillation (KD) can help models improve comprehensive capabilities without increasing size or data volume. However, most of the existing large model distillation techniques only consider applications on single-modal LLMs, or only use teachers to create new data environments for students. None of these methods takes into account the distillation of the most important cross-modal alignment knowledge in VLMs. We propose a method called Align-KD to guide the student model to learn the cross-modal matching that occurs at the shallow layer. The teacher also helps student learn the projection of vision token into text embedding space based on the focus of text. Under the guidance of Align-KD, the 1.7B MobileVLM V2 model can learn rich knowledge from the 7B teacher model with light design of training loss, and achieve an average score improvement of 2.0 across 6 benchmarks under two training subsets respectively.

Weakly-supervised Temporal Action Localization (WSTAL) aims to localize actions in untrimmed videos using only video-level supervision. Latest method introduce a pseudo label learning framework to bridge the gap between classification-based training and inference targets at localization. Typically, this framework employs a classification-based teacher model to generate pseudo labels, which are then used to train a regression-based student model for precise boundary prediction. However, the quality of these pseudo labels—critical to the student model’s performance—has not been systematically investigated, leading to suboptimal localization accuracy. In this paper, we propose a set of simple yet efficient mechanisms for pseudo label quality enhancement to build our FuSTAL framework. Unlike previous one or two stages methods, FuSTAL decomposes the learning process into three stages and enhances pseudo label quality at each one: cross-video contrastive learning for more informative initiative pseudo labels at theGeneration-Stage, prior-based filtering to remove the false positive proposals at theSelection-Stageand EMA-based distillation for smoother pseudo labels at theTraining-Stage. These designs supplement each other, and enhance action proposals’ quality with respect to the accuracy, true positive rate and smoothness. With the help of these comprehensive designs at all three stages, FuSTAL achieves an average mAP of 50.8% on the benchmark data THUMOS’14, outperforming the previous best method by 1.2%.

Semi-supervised Learning (SSL) reduces the need for extensive annotations in deep learning, but the more realistic challenge of imbalanced data distribution in SSL remains largely unexplored. In Class Imbalanced Semi-supervised Learning (CISSL), the bias introduced by unreliable pseudo-labels can be exacerbated by imbalanced data distributions. Most existing methods address this issue at instance-level through reweighting or resampling, but the performance is heavily limited by their reliance on biased backbone representation. Some other methods do perform feature-level adjustments like feature blending but might introduce unfavorable noise. In this paper, we discuss the bonus of a more balanced feature distribution for the CISSL problem, and further propose a Balanced Feature-Level Contrastive Learning method (BaCon). Our method directly regularizes the distribution of instances' representations in a well-designed contrastive manner. Specifically, class-wise feature centers are computed as the positive anchors, while negative anchors are selected by a straightforward yet effective mechanism. A distribution-related temperature adjustment is leveraged to control the class-wise contrastive degrees dynamically. Our method demonstrates its effectiveness through comprehensive experiments on the CIFAR10-LT, CIFAR100-LT, STL10-LT, and SVHN-LT datasets across various settings. For example, BaCon surpasses instance-level method FixMatch-based ABC on CIFAR10-LT with a 1.21% accuracy improvement, and outperforms state-of-the-art feature-level method CoSSL on CIFAR100-LT with a 0.63% accuracy improvement. When encountering more extreme imbalance degree, BaCon also shows better robustness than other methods.

In this paper, we propose HESIC, an end-to-end trainable deep network for stereo image compression (SIC). To fully explore the mutual information across two stereo images, we use a deep regression model to estimate the homography matrix, i.e., H matrix. Then, the left image is spatially transformed by the H matrix, and only the residual information between the left and right images is encoded to save bitrates. A two-branch auto-encoder architecture is adopted in HESIC, corresponding to the left and right images, respectively. For entropy coding, we use two conditional stereo entropy models, i.e., Gaussian mixture model (GMM) based and context based entropy models, to fully explore the correlation between the two images to reduce the coding bit-rates. In decoding, a cross quality enhancement module is proposed to enhance the image quality based on inverse H matrix. Experimental results show that our HESIC outperforms state-of-the-art SIC methods on InStereo2K and KITTI datasets both quantitatively and qualitatively. Code is available at https://github.com/ywz978020607/HESIC.