Reproduce “Zero-OVCD: Bridging Training-Free Foundation Models and Pseudo-Label Learning for Open-Vocabulary Change Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Daifeng Peng, Yuanke Peng, Haiyan Guan

A two-stage framework, Zero-OVCD, leverages vision foundation models to generate high-quality change pseudo-labels and subsequently trains a noise-aware change detector, addressing open-vocabulary change detection without requiring target-domain pixel-level annotations. The method demonstrated F1 score gains of 2.40 to 7.48 percentage points over its initial pseudo-labeling stage and up to 9.18 percentage points over prior training-free methods on binary change detection benchmarks.

Key insights

  • Integrating complementary mask sources and applying instance merging suppression effectively improves the coverage and quality of initial change candidate masks, reducing common segmentation inaccuracies.
  • Multiscale semantic verification, coupled with margin-based calibration, significantly enhances the reliability of semantic assignments by better discriminating between target and background categories and reducing false detections.
  • Training a dedicated change detector with noise-aware pseudo-label learning, which uses checkpoint voting and high-agreement sample selection, bridges training-free inference with the benefits of supervised learning, leading to more accurate and spatially coherent predictions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.