Large pre-trained vision-language models (VLMs) like CLIP have shown great potential for solving the unsupervised domain adaptation (UDA) problem. Existing prompt learning for UDA based on the unsupervised-trained VLMs requires distribution alignment between source and target domains in the common space for both vision and language branches. However, it is difficult for rough cross-domain alignment to maintain the discriminative semantic structure of both domains. Besides, the coarse features with non-informative noises due to ignoring the pseudo-label noises may cause failures to concentrate on precise semantics alignment. In this work, we propose a Prompt-Based Invertible Mapping Alignment (PIMA) method to incorporate discriminative domain knowledge into prompt learning, which is featured with refined cross-domain alignment in two separate space with a well-kept structure. Specifically, we design an invertible neural network-based homeomorphism mapping, and then achieve distribution alignment through such invertible mapping for connecting source and target visual feature space, which can preserve the data semantic structure. For better semantic alignment in vision-language space, we develop cross-modal implicit contrastive learning module to regularize non-informative features, which aims to find the low-rankness of implicit representation space. We conducted extensive experiments on three benchmark datasets to prove the advantages of our proposed PIMA over state-of-the-art methods.
8 November 2025 in Computer Vision and Image Understanding
Video coding plays a critical role for efficient transmission in surveillance camera sensors. Although long-term reference (LTR) has been fully studied in traditional hand-designed video coding approaches, its potential in learned video coding is still unexplored due to the highly unequal importance between long and short motion and the excessive motion overhead, especially for dense motion representation, e.g., optical flow. In this paper, we build an LTR baseline for learned surveillance video coding and propose an adaptive long–short modeling approach to address the above problem. Specifically, we first introduce LTR and propose a long–short context mining module to the authorized end-to-end video coding exploration model (EEM) from China’s AVS as a baseline. Since the quality of LTR significantly impacts its performance and importance, it is subsequently enhanced. Then, we propose a long–short motion adapter to address the unequal importance. Finally a historical motion guidance module is introduced to aid the motion decoding. Experimental results demonstrate that the proposed approach improves from a 1.86% BD-rate loss on EEM-4.1 to 13.89% BD-rate savings in YUV-PSNR compared with the anchor H.266/VVC under a low-delay P configuration. Although the current results are not comparable to the 44.01% gains of DCVC-FM, the proposed approach consumes less computational resources and we believe that integrating the proposed LTR method with stronger baselines will further boost the performance.
Referring expression segmentation (RES) aims to generate precise masks in images based on natural language descriptions. Existing approaches typically integrate text prompt into the segment anything model (SAM) via multi-modal encoder. However, these methods overlook a critical aspect: SAM’s inherently strength in processing visual prompt. To address this limitation, we propose a vision-language prompting framework that extends SAM to RES through a coarse-to-fine decoding strategy. Specifically, we first leverage existing multi-modal encoders to establish semantic consistency between visual and linguistic representations. Subsequently, a spatial–semantic interaction module generates a coarse target estimation that approximately localizes the referred regions. The resulting multi-modal spatial relations are then converted into visual prompts, enabling SAM to function as a vision-refined mask generator guided by visual information. In this framework, visual prompts provide the main guidance, while semantic prompts offer complementary support for fine-grained mask refinement. Extensive experiments demonstrate that our framework achieves performance on par with or surpassing state-of-the-art methods.
Efficient in-loop filtering is critical for the latest video-coding standard, versatile video coding (H.266/VVC). While parameter-adaptive mechanisms are effective in traditional adaptive loop filters and overfitted small-scale neural in-loop filtering, there is difficulty when deploying them in large-scale models since both schemes require explicit parameter signaling within the bitstream, leading to prohibitive overhead. Existing parameter-generation networks avoid these transmission costs but introduce an excessive number of parameters. To address this, we propose modeling convolutional parameters as a linear combination of pre-trained kernels, where weights are adaptively estimated via input-driven attention. Specifically, we propose a multi-scale parameter-adaptive convolution and its extension, with side information, enabling parameter adaptation without transmission overhead or significant computational costs. Furthermore, we have designed an efficient parameter adaptive in-loop filtering network with the proposed convolutions to balance parameter efficiency and reconstruction performance. To improve the distortion guidance provided by side information, we have incorporated gradient information. Experiments on VTM-11.0 demonstrate {7.89%, 18.25%, and 19.15%} bitrate savings for {Y, U, and V} components, outperforming fixed-parameter baselines by an average of 1.41% with negligible computational overhead.