TUNI: Unifying Pre-training and Fine-tuning with Modality-Aware Mutual Learning and Rectification for RGB-T Semantic Segmentation
Researchers have introduced TUNI, a novel framework designed to unify pre-training and fine-tuning for RGB-thermal (RGB-T) semantic segmentation. This approach aims to improve autonomous systems' perception in adverse conditions by addressing issues like suboptimal multi-modal feature extraction and fusion, and unbalanced modality dependency. TUNI incorporates a unified RGB-T encoder with a local module for emphasizing salient features across modalities, and employs modality-inverted contrastive mutual learning to mitigate bias during pre-training. For fine-tuning, modality rectification learning is used to leverage residual thermal information. AI
IMPACT This research could enhance the perception capabilities of autonomous systems in challenging environmental conditions.