MatchLive baseline β€” weights and test predictions

Files for MatchLive-baseline, the video-based baseline of the SoccerNet Challenge 2027 – Soccer Streaming VQA. Laid out like the code repository, so the weights download into place from its root:

hf download Becomebright/MatchLive-baseline --include "checkpoints/*" --local-dir .
Path Size What it is
checkpoints/rfdetr-l.onnx 122 MB RF-DETR-L detector (ball, goalkeeper, referee, player), ONNX
checkpoints/jersey_dinov3_convnext_tiny-20_matches.onnx 112 MB jersey-number classifier (DINOv3 ConvNeXt-tiny backbone + linear head; numbers 0–99 and "unreadable"), ONNX
results/submission_test.json 0.4 MB MatchLive (gemini-3.8-flash): predictions on the 700 test questions, in the challenge's submission format
results/submission_test_qwen.json 0.4 MB MatchLive (Qwen3.8-27B served locally): the same

Accuracy (%), scored with the challenge's codes/eval_predictions.py:

counting recall_event_time recall_event_player event_team event_player causal reasoning Average
submission_test.json 58.3 82.0 96.1 96.6 96.9 76.9 84.5
submission_test_qwen.json 42.9 70.5 79.1 84.7 72.3 53.8 67.2

License

The jersey-number classifier is built on a DINOv3 backbone and is subject to Meta's DINOv3 License. The predictions refer to the SoccerNet SN-VQA-2027 test questions, provided by SoccerNet under its data agreement.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support