LibreRTMDetl-seg
RTMDet-Ins-l COCO instance segmenter, repackaged for the LibreYOLO framework.
Source
Derived from https://github.com/open-mmlab/mmdetection at commit cfd5d3a985b0249de009b67d04f37263e11cdf3d and upstream checkpoint:
https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet-ins_l_8xb32-300e_coco/rtmdet-ins_l_8xb32-300e_coco_20221124_103237-78d1d652.pth
(SHA-256 78d1d6525c3065c3cf4f7033326f8a38ac6252fad28a7ff5f12e9db2d1a92b6a).
Copyright (c) 2018-2023 OpenMMLab. Licensed under the Apache License, Version 2.0.
Modifications
EMA weights were selected from the upstream checkpoint. data_preprocessor.* and batch-tracking buffers were omitted, bbox_head. keys were renamed to head., and the loaded state dict was saved with LibreYOLO checkpoint metadata schema v1.0 (task=segment). Learned model parameters are otherwise preserved.
Validation
Evaluated with LibreYOLO on full COCO val2017 (5000 images) at imgsz=640, conf=0.001, next to the official mmdetection references:
| Metric | LibreYOLO | Official |
|---|---|---|
| COCO val2017 mask mAP50-95 | 0.4368 | 43.7 |
| COCO val2017 box mAP50-95 | 0.5119 | 51.2 |
| SHA256 | 2df3d44206b14c994cd86eca9a75c73a9dbb9acd00293f158dbf1d7da3967310 |
Usage
from libreyolo import LibreYOLO
model = LibreYOLO("LibreRTMDetl-seg.pt")
res = model.predict("image.jpg")
res.masks # instance masks
res.boxes # boxes, scores, classes
License
Apache License 2.0. See the LICENSE and NOTICE files in this repository.