Long Cui

Long Cui, Weiyun Wang, Jie Shao et al.

Existing Multimodal Large Language Models (MLLMs) suffer from increased inference costs due to the additional vision tokens introduced by image inputs. In this work, we propose Visual Consistency Learning (ViCO), a novel training algorithm that enables the model to represent images of varying semantic complexities using different numbers of vision tokens. The key idea behind our method is to employ multiple MLP connectors, each with a different image compression ratio, to downsample the vision tokens based on the semantic complexity of the image. During training, we minimize the KL divergence between the responses conditioned on different MLP connectors. At inference time, we introduce an image router, termed Visual Resolution Router (ViR), that automatically selects the appropriate compression rate for each image patch. Compared with existing dynamic high-resolution strategies, which adjust the number of visual tokens based on image resolutions, our method dynamically adapts the number of visual tokens according to semantic complexity. Experimental results demonstrate that our method can reduce the number of vision tokens by up to 50% while maintaining the model's perception, reasoning, and OCR capabilities. We hope this work will contribute to the development of more efficient MLLMs. The code and models will be released to facilitate future research.

Wenpeng Li, Yongli Sun, Jinjun Wang et al.

Jointly utilizing global and local features to improve model accuracy is becoming a popular approach for the person re-identification (ReID) problem, because previous works using global features alone have very limited capacity at extracting discriminative local patterns in the obtained feature representation. Existing works that attempt to collect local patterns either explicitly slice the global feature into several local pieces in a handcrafted way, or apply the attention mechanism to implicitly infer the importance of different local regions. In this paper, we show that by explicitly learning the importance of small local parts and part combinations, we can further improve the final feature representation for Re-ID. Specifically, we first separate the global feature into multiple local slices at different scale with a proposed multi-branch structure. Then we introduce the Collaborative Attention Network (CAN) to automatically learn the combination of features from adjacent slices. In this way, the combination keeps the intrinsic relation between adjacent features across local regions and scales, without losing information by partitioning the global features. Experiment results on several widely-used public datasets including Market-1501, DukeMTMC-ReID and CUHK03 prove that the proposed method outperforms many existing state-of-the-art methods.