Multi-granularity 3D Visual Language Tracking from monocular videos: A benchmark and model

Abstract

Object tracking, a fundamental task in computer vision, has recently evolved beyond purely visual cues to incorporate natural language as high-level semantic guidance. Monocular 3D Visual Language Tracking (Mono3DVLT) extends this paradigm into the 3D domain, seeking to localize and track a referred 3D object in monocular video under natural language guidance. Despite its potential, existing approaches remain constrained by single-granularity language descriptions and weak temporal reasoning. Motivated by the diversity of natural instructions in real-world traffic scenes, we introduce a comprehensive framework for evaluating whether monocular 3D trackers perform under controlled variations i..