Tuning-free Universally-Supervised Semantic Segmentation
This work addresses semantic segmentation for computer vision applications, offering a universally applicable method that is incremental in improving alignment and robustness.
The paper tackles the problem of semantic segmentation under various supervision types by proposing a tuning-free framework that combines SAM masks and CLIP classification, addressing misalignment with a discrimination-bias aligned CLIP and a global-local consistent classifier. It achieves state-of-the-art or competitive performance across multiple datasets.
This work presents a tuning-free semantic segmentation framework based on classifying SAM masks by CLIP, which is universally applicable to various types of supervision. Initially, we utilize CLIP's zero-shot classification ability to generate pseudo-labels or perform open-vocabulary segmentation. However, the misalignment between mask and CLIP text embeddings leads to suboptimal results. To address this issue, we propose discrimination-bias aligned CLIP to closely align mask and text embedding, offering an overhead-free performance gain. We then construct a global-local consistent classifier to classify SAM masks, which reveals the intrinsic structure of high-quality embeddings produced by DBA-CLIP and demonstrates robustness against noisy pseudo-labels. Extensive experiments validate the efficiency and effectiveness of our method, and we achieve state-of-the-art (SOTA) or competitive performance across various datasets and supervision types.