World Models as Data Engines: Annotation-Preserving Synthetic Video for Soccer Athlete Re-Identification

Re-identifying athletes across frames and cameras underpins player tracking and game-state reconstruction in soccer, yet training robust re-identification (ReID) models requires large volumes of identity-labeled footage that are costly and slow to annotate. Synthetic data is an attractive remedy, but game-engine renders exhibit a wide appearance gap to broadcast video, while photorealistic generative models produce frames whose player boxes and identities must be re-annotated from scratch. This work shows that a world foundation model can act as an annotation-preserving data engine: because generation is conditioned on spatial structure aligned with a source clip, the source’s bounding boxes and track identities remain approximately valid on the generated video. NVIDIA Cosmos 3 is first domain-specialized to soccer broadcast via LoRA-based supervised fine-tuning on clips from the SoccerNet Game State Reconstruction dataset, prompted with structured captions from Qwen3-VL. Multi-control video-to-video transfer (edge, depth, segmentation) is then applied to SportsMOT soccer clips, re-rendering each under diverse lighting conditions and jersey colors while preserving player layout. Merging the generated variants into ReID training – with evaluation on real, held-out SportsMOT data only across two different architectures – improves mAP by 7.60 ± 1.90 and rank-1 accuracy by 7.85 ± 1.75 percentage points over a real-only baseline, with joint augmentation outperforming synthetic-only pre-training.

Hali tarjima qilinmagan