convnextv2-soccer-play-classifier

ConvNeXtV2 Soccer Play-State Classifier

This model classifies a soccer broadcast frame as active play or out of play. It was locally fine-tuned from timm/convnextv2_base.fcmae_ft_in22k_in1k.

Model details

• Architecture: ConvNeXtV2 Base
• Task: binary image classification
• Resolution: 224 × 224 RGB
• Classes: 0 = out_of_play, 1 = in_play
• Selected checkpoint: best local checkpoint
• Best validation accuracy: 91.87%
• Best epoch: 14
• Base model license: CC BY-NC 4.0

The metric comes from a local validation split and has not been independently reproduced on a public benchmark.

Intended use

Non-commercial research and prototyping for soccer broadcast segmentation, play-state filtering and editing assistance. Commercial use is prohibited by the inherited CC BY-NC 4.0 terms unless the necessary additional permissions are obtained.

Training procedure

The original classification head was replaced with a two-class head. The model used staged backbone freezing/unfreezing, AdamW and cosine scheduling.

• Batch size: 32
• Initial learning rate: 1e-4
• Weight decay: 1e-4
• Image size: 224
• Normalization mean: [0.485, 0.456, 0.406]
• Normalization standard deviation: [0.229, 0.224, 0.225]

Training-data disclosure

Training images are not distributed. They were locally extracted from soccer broadcast footage and grouped into out_of_play and in_play. Some source footage was collected from publicly accessible Bilibili videos, but public availability does not establish redistribution permission. No videos, frames, audio, subtitles, uploader information or platform metadata are included.

The source material has not undergone complete work-by-work copyright clearance. Users must assess their intended use under applicable copyright, privacy, publicity and platform rules. Rights holders may request review or removal through the Hugging Face repository contact/discussion channel.

Usage

pip install -r requirements.txt
python inference.py path/to/frame.jpg --checkpoint model.pth

The checkpoint contains model_state_dict and training metadata. Load pickle-based PyTorch files only from trusted sources.

Limitations

• Still frames lack temporal context and can be ambiguous.
• Replays, graphics, close-ups and uncommon camera angles may reduce accuracy.
• Generalization outside the local broadcast distribution is unverified.
• The model may learn broadcaster-specific visual features.
• Do not use it for surveillance or identity-related decisions.

License

Creative Commons Attribution-NonCommercial 4.0 International. Credit the timm base model and original ConvNeXtV2 work. See LICENSE.md.