rtmdet-m-640
RTMDet-M (MMDetection) fine-tuned on the basketball dataset at 640x640, the mmdet-default-warmup ("recipe warmup") training. A short-warmup ablation variant exists (not registered here) and was found to be statistically indistinguishable on the harness test slice.
Metrics
Measured on basketball-player-detection-3 (test split), via the object-detection-eval harness.
| Metric | 5-class | 10-class |
|---|---|---|
| mAP@50:95 | 0.628 | 0.580 |
| mAP@50 | 0.878 | 0.779 |
Preprocessing
- Resize:
letterbox(alignment:top_left) - Normalize:
mean_std - Channel order:
BGR - Input shape:
(3, 640, 640)(float32)
Provenance
- Source repo: https://github.com/open-mmlab/mmdetection
- Training config:
docs/provenance/configs/rtmdet_m/rtmdet_basketball.py - Hardware: vast.ai A100 80GB (single GPU)
- Command:
torchrun --nproc_per_node=1 tools/train.py docs/provenance/configs/rtmdet_m/rtmdet_basketball.py --launcher pytorch
Usage
This ONNX file is one of the 7-model roster benchmarked in object-detection-eval. Load it through the registry for verified, hash-checked download and the exact preprocessing recipe above:
from object_detection_eval.registry import ModelRegistry, download_weights
registry = ModelRegistry.from_directory("registry")
card = registry.get("rtmdet-m-640")
weights_path = download_weights(card)