Geometry-First Pipeline Outperforms Depth Estimation in Multi-Camera 3D Tracking

A new arXiv paper (2608.07579) presents a geometry-first approach to multi-camera 3D tracking that significantly outperforms depth-estimation methods in a Sim2Real setting. The study, conducted for the AI City Challenge 2026 Track 1, focuses on large indoor warehouses where depth data is available only during training and validation, forcing RGB-only inference. The authors tested two RGB-only pipelines: a geometry-first method using YOLO11x detection, homography lifting, class-level 3D size priors, multi-camera fusion, and offline tracklet stitching; and a pseudo-LiDAR approach that back-projects monocular depth (D4RT, Metric3D v2) into a fused point cloud for a 3D detector (V-DETR). The geometry-first pipeline achieved a 3D HOTA of 13.0 and LocA of 51.6, while the pseudo-LiDAR method collapsed to 0.12 HOTA and 9.2 LocA. The authors attribute the success to cross-view geometric consistency rather than monocular depth accuracy. The paper is available on arXiv.