charlesjvt/rtmpose-l-h36m-384x288

RTMPose-l fine-tuned on Human3.6M (384x288)

RTMPose-l 2D keypoint detector fine-tuned on Human3.6M, used by Kineo for the Human3.6M benchmarks.

Files

• rtmpose_h36m.pth — checkpoint (state_dict + ema_state_dict)
• rtmpose-l_8xb256-420e_h36m-384x288.py — MMPose training/inference config

Details

Architecture	RTMPose-l (CSPNeXt-l + RTMCCHead, SimCC)
Input size	384x288
Keypoints	17 (COCO layout)
Dataset	Human36mCocoDataset, h36m_train_coco_10fps.json
Initialised from	rtmpose-l_simcc-aic-coco_pt-aic-coco_420e-384x288-97d6cb0f_20230228.pth
Epoch	104 (best coco/AP of that run; 420-epoch schedule stopped early)
coco/AP on h36m_test_coco_10fps.json	0.9667
Training run	rtmpose-l_8xb256-420e_h36m-384x288_20250409_090631, seed 21

Usage

from mmpose.apis import init_model

model = init_model(
    "rtmpose-l_8xb256-420e_h36m-384x288.py",
    "rtmpose_h36m.pth",
    device="cuda:0",
)

License

Research and evaluation use only; commercial use or redistribution requires prior written consent. Contact: guillaume.lavoue@enise.ec-lyon.fr

Human3.6M is subject to its own license — see http://vision.imar.ro/human3.6m/.