// the find
NVIDIA-AI-IOT/trt_pose
Real-time pose estimation accelerated with NVIDIA TensorRT
trt_pose is a human pose estimation library that runs keypoint models converted from PyTorch to TensorRT with torch2trt, aimed at NVIDIA Jetson boards. It ships pretrained ResNet18 and DenseNet121 models trained on MSCOCO, plus training scripts for any keypoint task in COCO format. It suits people who need a pose model on edge hardware, not general server inference.
- The README publishes FPS numbers per board: the ResNet18 224x224 model runs at 22 FPS on a Jetson Nano and 251 FPS on a Xavier. You can check whether it fits your hardware before installing anything.
- Keypoint grouping (peak finding, part-affinity connection, Hungarian matching via munkres) is written in C++ rather than Python. That is the step that would otherwise dominate CPU time on a Nano.
- Training is not limited to people. Any MSCOCO-format keypoint dataset works, and tasks/human_pose/experiments has configs across ResNet, DenseNet, DLA and MNASNet backbones to start from.
- The last push was August 2022. The install steps assume the older torch2trt build with plugins, and expect to fight the toolchain on a current JetPack.
- The weights are Google Drive links, not release assets or repo files. There is no versioning or checksum, so link rot is a real risk for a project this old.
- The entry point is a Jupyter notebook, and the README itself says you may need to edit it for your model. There is no CLI or packaged inference API, so anything beyond the demo means writing your own glue.