Title: CoRe-CLIP: Restoring Final-Layer Patch Coherence for Training-Free Open-Vocabulary Semantic Segmentation
-Journal/Conference: British Machine Vision Conference (BMVC) 2026
-Authors: Junghun Kim, Jongpil Jeong
-DOI: 
-Journal/Conference Link: https://bmvc2026.org/

Abstract: Vision-language models such as CLIP provide strong open-vocabulary recognition
ability, but dense prediction remains challenging because image-level contrastive training does not directly enforce spatially coherent patch representations. Recent trainingfree OVSS methods improve dense CLIP inference by correcting the final visual block
or modifying attention behavior, yet the features used for text matching can still be fragmented within an object. We observe that CLIP Layer-2 patch affinity provides a useful structural cue for reorganizing these representations. We therefore introduce CoReCLIP, a training-free feature re-aggregation module that converts this shallow affinity into a sparse local routing distribution and applies it to raw final-layer features before text projection. Layer 2 supplies the routing support and weights, while the final layer retains the semantic vectors matched with text. CoRe-CLIP preserves the base method’s attention mechanism, text encoder, class names, and evaluation protocol, while introducing no learnable parameters, external foundation model, or additional CLIP forward pass. Across eight benchmarks and four training-free OVSS base methods, including NACLIP, ResCLIP, SCLIP, and ClearCLIP, CoRe-CLIP improves almost all paired settings and increases the average performance of every base method. Diagnostic analyses further show that the gains do not come from generic local averaging alone and that the routing improves final-feature coherence.


-Status: Submitted (2026.05.29)