COAL: Counterfactual and Observation-Enhanced Alignment Learning for Discriminative Referring Multi-Object Tracking
COAL addresses a core problem in tracking multiple objects by combining explicit semantic knowledge with learning techniques to overcome data scarcity. It introduces Explicit Semantic Injection (ESI) from a VLM and Counterfactual Learning (CFL) via an LLM, integrated into a Hierarchical Multi-Stream Integration (HMSI) architecture. This resolves the conflict between needing high discrimination and having sparse supervision, leading to state-of-the-art tracking performance.