Apppublic

Tamoghna1995/rtdetr-sportsmot

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes

App README

RT-DETRv2 · SportsMOT player detection

Live demo for `smallTech/rtdetr-sportsmot` — a single-class player detector fine-tuned from `PekingU/rtdetr_v2_r50vd` on the SportsMOT train split, built as the detection stage of a ByteTrack tracking-by-detection pipeline.

Hit 🎲 Random test image to run detection on a random frame from the SportsMOT test split, or upload your own. Inference runs on ZeroGPU (a per-request slice of an NVIDIA GPU). Benchmarks — mAP@50 0.938 on 45 unseen val sequences — live on the model card.

The random example frames are served from an HF Storage Bucket (Tamoghna1995/rtdetr-sportsmot-examples) mounted read-only at /data. The model repo is gated; this Space reads a HF_TOKEN secret to download the weights.

Repo layout

—README.md (this file) — the Space card.
—app.py — the Gradio app.
—requirements.txt — Python deps (torch / gradio / spaces are platform-managed and intentionally unlisted).

In the workspace repository this Space is tracked as a git submodule at Tamoghna1995/spaces/rtdetr-sportsmot.