From road placement to vehicle labels
Explore four real artifacts from the UE5 capture workflow: road geometry, projected keypoints, rendered camera poses, and dataset labels.
The source example is available above; interactive controls load when JavaScript is enabled.
The research result
The published dataset contains 2,510 frames with a 24-point schema. In the completed matched 2D-transfer experiments, adding synthetic data reduced real-data OKS-mAP by 1.7 pp from the weak checkpoint and 9.5 pp from the strong one. The +2 pp criterion was not met.
The capture workflow and its transfer result are separate deliverables. Read the experiment or inspect the dataset.
Try live vehicle detection on your own image. That Space uses the separate main 14-point detector trained on real data, not the synthetic-only 24-point model.
The figures are captured/rendered artifacts from the published documentation. Epic asset files are excluded from the dataset; see provenance and terms.
Sources and reproduction
The explorer uses published artifacts from kiselyovd/ue5-vehicle-synth. Inspect the pinned source and the structured showcase data.