Imbalance-aware and direction-agnostic label assignment for RGBT small object detection

Label assignment is crucial for object detection training, yet small objects often receive few low-quality positive samples, causing model bias toward large objects. This paper proposes Imbalance-Aware Weighted Wasserstein Distance (IAWD) and direction-agnostic Euclidean Centerness (ECenterness) to improve positive sample quantity and quality for small objects. IAWD replaces IoU-based assignment by measuring anchor-GT similarity without requiring overlap, with weights determined by object scale and distribution distance imbalance. ECenterness linearly quantifies proximity via the shortest path, eliminating directional bias. Additionally, a Cross-modal feature fusion (CMFF) module is introduced for multi-modal prediction. Experiments on RGBT-Tiny and RGBT-Droneperson datasets validate the effectiveness of our approach.