RAGTrack
The official implementation of a CVPR 2026 paper on language-aware RGB-thermal object tracking, combining a multi-modal transformer encoder with retrieval-augmented context reasoning.
RAGTrack is a research codebase for RGB-thermal (RGBT) visual object tracking that adds a linguistic reasoning path on top of vision-only trackers, addressing weak appearance modeling, search redundancy and the gap between RGB and thermal modalities. It introduces a Multi-modal Transformer Encoder for joint visual-language modeling, an Adaptive Token Fusion module that dynamically selects target-relevant tokens and exchanges channels between modalities, and a Context-aware Reasoning Module that retrieves relevant context for each frame using retrieval-augmented generation.
The authors built four RGB-T-L benchmarks by adding textual descriptions to the existing GTOT, RGBT210, RGBT234 and LasHeR tracking datasets through an MLLM-based annotation pipeline. The repository includes training and evaluation code, pretrained models depending on CLIP and Qwen-VL, and published results and benchmark data.