rfdetr-basketball-player-ball-referee-detection

This model is a fine-tuned version of Roboflow/rf-detr-nano on an labeled basketball object detection dataset from Roboflow.

• Roboflow Dataset - https://universe.roboflow.com/roboflow-jvuqo/basketball-player-detection-3-ycjdo/dataset/18
• Same Dataset as above but exported to HF - https://huggingface.co/datasets/koppolusameer/basketball-coco-20260416

It achieves the following results on the evaluation set:

• Loss: 11.6183
• Map: 0.4681
• Map 50: 0.7416
• Map 75: 0.5072
• Map Small: 0.3566
• Map Medium: 0.6316
• Map Large: -1.0
• Mar 1: 0.1966
• Mar 10: 0.5356
• Mar 100: 0.5675
• Mar Small: 0.4527
• Mar Medium: 0.7196
• Mar Large: -1.0
• Map Ball: 0.1805
• Map Player: 0.6228
• Map Referee: 0.601

Model description

The model is the RF-DETR Nano Model (https://huggingface.co/Roboflow/rf-detr-seg-nano).

Intended uses & limitations

Model is finetuned to detect players, ball, and referee in basketball broadcast images. Below are a few limitations

• The ball is only a few pixels in size and often gets mixed with the background and confuses the model (as described by mAP results)
• Images are mostly from Celtics games and hence there may be a lack of training data diversity

Training and evaluation data

• Train, Validation, and Test Splits can be found here - https://huggingface.co/datasets/koppolusameer/basketball-coco-20260416

Training procedure

• Model was trained for 10 epochs and with the same train, valid, and test split as in the dataset
• Training data was randomly augmented before finalizing the train set, where the augmentations done on an image could include at least one of
  • Horizontal flipping
  • Brightness and Contrast changes
  • Saturation changes
  • Blurring
  • Shifting, scaling, and Rotation

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• num_epochs: 10

Training results

Framework versions

• Transformers 5.10.2
• Pytorch 2.11.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.2