Zhikang Liu

Zhikang Liu, Lanyun Zhu

Contemporary segmentation methods are usually based on deep fully convolutional networks (FCNs). However, the layer-by-layer convolutions with a growing receptive field is not good at capturing long-range contexts such as lane markers in the scene. In this paper, we address this issue by designing a distillation method that exploits label structure when training segmentation network. The intuition is that the ground-truth lane annotations themselves exhibit internal structure. We broadcast the structure hints throughout a teacher network, i.e., we train a teacher network that consumes a lane label map as input and attempts to replicate it as output. Then, the attention maps of the teacher network are adopted as supervisors of the student segmentation network. The teacher network, with label structure information embedded, knows distinctly where the convolution layers should pay visual attention into. The proposed method is named as Label-guided Attention Distillation (LGAD). It turns out that the student network learns significantly better with LGAD than when learning alone. As the teacher network is deprecated after training, our method do not increase the inference time. Note that LGAD can be easily incorporated in any lane segmentation network.

Junshan Hu, Jialiang Mao, Zhikang Liu et al.

Conventional Vision-Language Models(VLMs) typically utilize a fixed number of vision tokens, regardless of task complexity. This one-size-fits-all strategy introduces notable inefficiencies: using excessive tokens leads to unnecessary computational overhead in simpler tasks, whereas insufficient tokens compromise fine-grained visual comprehension in more complex contexts. To overcome these limitations, we present TokenFLEX, an innovative and adaptable vision-language framework that encodes images into a variable number of tokens for efficient integration with a Large Language Model (LLM). Our approach is underpinned by two pivotal innovations. Firstly, we present a novel training paradigm that enhances performance across varying numbers of vision tokens by stochastically modulating token counts during training. Secondly, we design a lightweight vision token projector incorporating an adaptive pooling layer and SwiGLU, allowing for flexible downsampling of vision tokens and adaptive selection of features tailored to specific token counts. Comprehensive experiments reveal that TokenFLEX consistently outperforms its fixed-token counterparts, achieving notable performance gains across various token counts enhancements of 1.6%, 1.0%, and 0.4% with 64, 144, and 256 tokens, respectively averaged over eight vision-language benchmarks. These results underscore TokenFLEX's remarkable flexibility while maintaining high-performance vision-language understanding.