Skip to content
Open access

Occlusion-aware test-time adaptation for robust traffic sign image-text matching

Oct 2026 · Discover Artificial Intelligence · Vol 6 · 0 citations · 35 references

Abstract

Traffic-sign image–text matching supports retrieval-based vision–language perception in driving scenes, but partial occlusion can hide the visual information needed to link a traffic sign to its description. The robustness can be enhanced by adding experimental data or retraining the model on corrupted image–text pairs. Both of these would need extra data, and the corruptions in the training data may not be the same as those at test time. We evaluate OATA, a source-free test-time adaptation pipeline that combines BatchNorm TENT with a lightweight image–text pair reranker. Given unlabeled occluded test images, TENT updates only the BatchNorm scale and shift parameters while the text encoder remains frozen. The reranker then uses image–text pair scores to reorder the Top-5 candidates. Under a dataset-level protocol with 1,512 images and 3,024 captions, full OATA achieves 96.69% image-to-text Recall@1 and 96.43% text-to-image Recall@1. These results suggest that a combination of source-free BatchNorm adaptation with lightweight pair reranking can improve the top-rank performance of traffic-sign image–text retrieval on controlled synthetic occlusion shifts. This study also provides an evaluation protocol for future robustness studies.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.