OpenLongTail: Generative Scaling of Long-Tail Driving Data
OpenLongTail turns monocular and otherwise incomplete long-tail driving videos into multi-view training assets.
The paper says scarce edge-case data is a bottleneck for robust autonomous-driving policies, especially when real-world clips lack full sensor coverage. OpenLongTail uses pose-informed extrapolative view synthesis to generate missing views from heterogeneous sources. It adds Plucker ray geometry to improve cross-view consistency and temporal alignment. The authors report improved closed-loop robustness on long-tail events, along with gains in visual fidelity, consistency, and ego-trajectory recovery. HF Daily Papers' note
The paper says scarce edge-case data is a bottleneck for robust autonomous-driving policies, especially when real-world clips lack full sensor coverage. OpenLongTail uses pose-informed extrapolative view synthesis to generate missing views from heterogeneous sources. It adds Plucker ray geometry to improve cross-view consistency and temporal alignment. The authors report improved closed-loop robustness on long-tail events, along with gains in visual fidelity, consistency, and ego-trajectory recovery. HF Daily Papers' note
score 5