DETRPose-S-CROWDPOSE

DETRPose-S-CROWDPOSE is a real-time object detection model introduced in the paper DETRPose: Real-Time End-to-End Multi-Person Pose Estimation via Modified Transformer Decoder and Novel Denoising Keypoints.

Repository: https://github.com/SebastianJanampa/DETRPose
Paper: https://huggingface.co/papers/2506.13027

📝 Model Description

DETRPose introduces the first real-time end-to-end framework for multi-person pose estimation. By leveraging the hybrid encoder from RT-DETR and the lightweight decoder architecture of D-FINE, DETRPose achieves low-latency inference without sacrificing accuracy. The model introduces two primary methodological advancements:

Pose-LQE Layer: A specialized head designed to improve confidence scores.
Advanced Training Paradigm: Incorporates Denoising Keypoints and a custom Keypoint Similarity Varifocal loss function, ensuring robust learning and superior localization performance.

Model	Dataset	AP	#Params	Latency	GFLOPs
DETRPose-S	CROWDPOSE	67.4	11.5 M	4.80 ms	31.3

🚀 Installation

To use this model, you need to install the inference-ready branch of the DETRPose repository. You can directly install the inference-ready branch using pip:

pip install git+https://github.com/SebastianJanampa/DETRPose.git@inference_only

💻 Usage

This branch is designed to be easy to use for inference. Here is a quick example of how to load a model and run it on a live webcam feed.

from detrpose import DETR

# Initialization
model = DETR(model='detrpose_hgnetv2_s_crowdpose')

# Inference
model(source=0) # inference on a webcam

📜 Citation

If you use DETRPose or its methods in your work, please cite the following BibTeX entries:

@misc{janampa2025detrpose,
      title={DETRPose: Real-time end-to-end transformer model for multi-person pose estimation}, 
      author={Sebastian Janampa and Marios Pattichis},
      year={2025},
      eprint={2506.13027},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.13027}, 
}