Tuning-free Universally-Supervised Semantic Segmentation

TL;DR

This paper introduces a tuning-free semantic segmentation framework that leverages CLIP's zero-shot capabilities and a novel discrimination-bias aligned CLIP to improve mask classification, achieving state-of-the-art results without extensive tuning.

Contribution

The work proposes a universal, tuning-free segmentation method using DBA-CLIP and a global-local classifier, addressing alignment issues and robustness to noise.

Findings

Achieves SOTA or competitive results across multiple datasets.

Demonstrates robustness against noisy pseudo-labels.

Provides an efficient, tuning-free segmentation framework.

Abstract

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.