RTMDet-Ins s · person-only, lite (ONNX)

A trimmed RTMDet-Ins s model that returns only people, for in-browser use with ONNX Runtime Web (WebGPU). It is used by the VEC video editor to find people in video frames on the user's device.

Input: input, static 1x3x640x640, BGR, aspect-preserving resize, pad [114, 114, 114] (same as the original). Outputs: dets [1, n, 5] (x1, y1, x2, y2, score), labels [1, n] (always 0 = person), masks [1, n, 80, 80] mask logits at 1/8 of the input (apply sigmoid and upscale yourself), n <= 20.

Origin

• Converted ONNX: tori29umai/rtmdet-ins-onnx-person-masks, models/s/rtmdet-ins_s_640x640.onnx, revision e5218e8d47d7c50077f27f9893a28e0ef2ad5656 (Apache-2.0).
• Weights: OpenMMLab MMDetection RTMDet-Ins s, COCO, rtmdet-ins_s_8xb32-300e_coco_20221121_212604-fdc5d7ec.pth (MMDetection, Apache-2.0), exported with MMDeploy v1.3.1.
• Training data: COCO 2017. Annotations © COCO Consortium, CC BY 4.0. Images belong to their Flickr owners and are not included here.

Modifications (changed files: both .onnx)

Made on 2026-10-31 by editing the ONNX graph with onnx (script: recortar-rtmdet.py):

1. Only class 0 (person) is scored before TopK and NMS.
2. Fewer candidates: pre-NMS TopK 5000 → 1000; NMS boxes per class 200 → 20; keep_top_k 100 → 20.
3. NMS score threshold 0.05 → 0.2; IoU threshold 0.5 → 0.6.
4. The final mask Resize (to 640x640) and Sigmoid were removed: masks are 80x80 logits.
5. The fp16 file was converted with onnxconverter-common (keep_io_types=True; Cast nodes and backbone/stage2 left in float32). No retraining: the weights are those of the original model (rounded to float16 in the fp16 file).

License and attribution

Apache License 2.0; see LICENSE and NOTICE. Copyright 2018-2023 OpenMMLab. ONNX conversion by tori29umai. Modifications by OscarEspana. Provided "AS IS", without warranties. Not affiliated with or endorsed by OpenMMLab or tori29umai.