Unveiling the Knowledge of CLIP for Training-Free Open-Vocabulary Semantic Segmentation

Training-free open-vocabulary semantic segmentation aims to explore the potential of frozen vision-language models (VLM) for segmentation tasks. Recent works reform the inference process of CLIP and utilize the features from the final layer to reconstruct dense representations for segmentation, demonstrating promising performance. However, the final layer tends to prioritize global components over local representations, leading to suboptimal robustness and effectiveness of existing methods. In this paper, we propose CLIPSeg, a novel training-free framework that fully exploits the diverse knowledge across layers in CLIP for dense predictions. Our study unveils two key discoveries: Firstly, the features in the middle layers exhibit high locality awareness and feature coherence compared to the final layer, based on which we propose the coherence enhanced residual attention module that generates semantic-aware attention. Secondly, despite not being directly aligned with the text, the deep layers capture valid local semantics that complement those in the final layer. Leveraging this insight, we introduce the deep semantic integration module to boost the patch semantics in the final block. Experiments conducted on 9 segmentation benchmarks with various CLIP models demonstrate that CLIPSeg consistently outperforms all training-free methods by substantial margins, e.g., a 7.8 % improvement in average mIoU for CLIP with a ViT-L backbone, and competes with learning-based counterparts in generalizing to novel concepts in an efficient way.