What it does

tr_t_pose takes pre-trained human pose estimation models and accelerates them through NVIDIA TensorRT so they run in real time on Jetson devices. It detects standard COCO keypoints—eyes, elbows, ankles, the full skeleton—using a webcam or camera input. The project also includes training scripts for any keypoint task in MSCOCO format, so you’re not locked into human pose.

The interesting bit

The speedup is substantial: a ResNet-18 variant hits 22 FPS on a Jetson Nano and 251 FPS on Jetson Xavier. That’s not magic; it’s torch2trt doing the heavy lifting to convert PyTorch graphs to TensorRT engines. The project is essentially a well-packaged integration of existing pieces—OpenPose-style part affinity fields, Simple Baselines architecture, and NVIDIA’s own converter—rather than a novel model.

Key highlights

• Two pre-trained models provided: ResNet-18 (224×224) and DenseNet-121 (256×256)
• Live Jupyter notebook demo with camera input included
• Training pipeline supports any MSCOCO-format keypoint dataset
• Weights are ~80–84MB; hosted on Google Drive
• Companion project trt_pose_hand extends this to hand/gesture recognition

Caveats

• Setup requires manual dependency installation including torch2trt with plugins; no pip package
• Model weights live on Google Drive, not in the repo
• Notebook may need modification depending on which model you choose

Verdict

Grab this if you’re building a Jetson-based robot, kiosk, or fitness app and need pose detection without a GPU server in the loop. Skip it if you’re on non-NVIDIA hardware or need multi-person tracking with heavy occlusion—this is optimized for the edge, not accuracy benchmarks.