This repository contains RTMDet-Ins COCO instance segmentation models converted to ONNX for ONNX Runtime.
The intended use is lightweight person bounding box and person mask extraction before downstream pose estimation. The models output COCO classes; the sample code filters class 0 (person) and applies an additional mask-overlap NMS step to remove duplicate person detections.
Files
Variant
File
Size
Notes
tiny
models/tiny/rtmdet-ins_tiny_640x640.onnx
~23 MB
Fastest, lower confidence on crowded scenes
s
models/s/rtmdet-ins_s_640x640.onnx
~41 MB
Recommended default
m
models/m/rtmdet-ins_m_640x640.onnx
~110 MB
Stronger than s, slower
l
models/l/rtmdet-ins_l_640x640.onnx
~228 MB
Can over-detect on crowded anime/group images
x
models/x/rtmdet-ins_x_640x640.onnx
~406 MB
Highest quality among this set, slowest
Each model folder also includes the MMDeploy deploy.json, pipeline.json, and detail.json generated during conversion.
Recommended Settings
For person mask extraction:
Variant
Recommended score_thr
mask_overlap_thr
Notes
tiny
0.30
0.85
Use when speed matters
s
0.30
0.85
Recommended default
m
0.30
0.85
Good balance if s misses people
l
0.40
0.85
Higher threshold helps reduce over-detection
x
0.30
0.85
Best quality, heavier model
On the bundled conversion test image, the expected count was five people. s kept five people with score_thr=0.30 and mask_overlap_thr=0.85. Without mask-overlap NMS, one person could be detected twice.
The script writes a visualization image and prints detected person boxes as JSON.
Input / Output
Input:
Static 1x3x640x640
BGR normalization
Resize with aspect ratio preserved
Pad with [114, 114, 114]
Raw ONNX outputs:
dets: [1, N, 5] (x1, y1, x2, y2, score)
labels: [1, N]
masks: [1, N, 640, 640] or [N, 640, 640]
License
These models are converted from OpenMMLab MMDetection RTMDet-Ins COCO checkpoints. The upstream project is licensed under Apache License 2.0. See LICENSE and NOTICE.md.