Retrainable RTMO-s Model
This repository provides a fully retrainable RTMO-s checkpoint for 2D human pose estimation in the MMPOSE framework.
The RTMO one-stage model family (variants T, S, M, L) was originally released by the OpenMMLab team with pre-trained weights, but those official PyTorch .pth checkpoints do not preserve all parameter keys—making fine-tuning or re-training impossible within MMPOSE (see discussion #3076).
To address this limitation, PESI has faithfully reproduced the RTMO-s training procedure on the MS COCO 2017 dataset using the exact configuration from the official MMPOSE RTMO project. Our checkpoint preserves every model key, enabling you to fine-tune on custom datasets (e.g. Body7, MPII) or continue training from this strong baseline.
- Original implementation & configs: MMPOSE RTMO project
- Issue on missing keys: MMPOSE #3076
Key Features
- One-stage RTMO-s architecture as described in the MMPOSE RTMO project
- All keys retained in the PyTorch checkpoint for full retrainability (unlike the official weights)
References
- OpenMMLab MMPOSE RTMO project: “RTMO: Real-time One-Stage Multi-person Pose Estimation”
- Issue discussion on missing keys in official RTMO checkpoints