runrazor detector models
Downloadable subject-detector weights for runrazor.
These are stock YOLOX object detectors, exported to ONNX and used at runtime
restricted to the COCO person class (skiers are detected as people; runrazor's
tracking layer handles subject selection).
runrazor fetches these on demand: rrreframe clip.mp4 --detector nn --nn-model yolox_s
downloads the file once, verifies its SHA-256, and caches it. The smaller
yolox_nano ships inside the runrazor package itself and needs no download.
Files
| file | base model | input | notes |
|---|---|---|---|
yolox_s-v1.onnx |
YOLOX-S | 640 | best accuracy/speed balance; recommended upgrade |
yolox_m-v1.onnx |
YOLOX-M | 640 | highest recall; ~2.4x slower than S on CPU |
Each file is pre-stamped with its runrazor runtime settings (format=yolox_raw,
conf, class_id=0) in the ONNX metadata, so it is a drop-in -- no extra flags.
Attribution and license
These weights are derived from YOLOX (https://github.com/Megvii-BaseDetection/YOLOX),
Copyright (c) 2021-2022 Megvii Inc., licensed under the Apache License 2.0
(see LICENSE). The networks were not retrained; the upstream .pth weights were
exported to ONNX. YOLOX is pretrained on the COCO dataset.