pose_inference

pose_inference is a C++ library and set of command-line tools for batched keypoint inference on synchronized image and video inputs. In this workspace it is usually used after a detection stage has already produced person crops or bounding boxes.

The code assumes a fixed batch size at configure time. In practice that usually means one input stream per camera.

Dependencies

This project depends on:

• OpenCV with CUDA support
• TensorRT 10
• Eigen3
• cpp_utils
• tensorrt-cpp-api

The simplest way to satisfy those dependencies is to use the container from the root repository:

• Docker-OpenCV-TensorRT-Dev

Build

Set any required environment variables first. The important ones for this repo are:

export OPENCV_VERSION=4.13.0
export BATCH_SIZE=5

BATCH_SIZE is compiled into the binaries. The ONNX model and configuration need to match it.

Build and install:

sudo ./build_install.sh

Configuration

• Default config file: cfg/pose_all_config.json
• Optional config path can be passed to most executables as the third argument
• The model should be exported with a fixed batch size that matches BATCH_SIZE
• The current implementation is structured around RTMPose-style models

Executables

inference_benchmark

./build/inference_benchmark <input_dir> <output_dir> [config_path]

Reads BATCH_SIZE .jpg files from input_dir, runs keypoint inference, and writes images with keypoints drawn to output_dir.

video_inference_export

./build/video_inference_export <input_dir> <output_dir> [config_path]

Reads BATCH_SIZE synchronized .mp4 files from input_dir and writes one JSON keypoint log per input video to output_dir.

keypoint_overlay

./build/keypoint_overlay <input_dir> <output_dir> [config_path]

Reads .mp4 files and matching keypoint JSON files from input_dir, then writes overlay videos to output_dir. It also writes a tiled 4cams.mp4 summary video. The optional config_path argument is accepted for interface consistency but ignored by this executable.

Tested Configurations

• Intel Xeon W-2145, RTX 2080 Super, Ubuntu 20.04, CUDA 11.8, TensorRT 8.6.1.6, OpenCV 4.10.0, BATCH_SIZE=5 Preprocess about 1 ms, inference about 4 ms, postprocess about 1 ms over 1000 samples
• Intel Core i7-12700H, RTX 3050 Ti, Ubuntu 24.04, CUDA 12.9, TensorRT 10.14.1.48, OpenCV 4.13.0, BATCH_SIZE=5 Preprocess above 3 ms, inference about 11 ms, postprocess above 1 ms over 1000 samples
• AMD Ryzen 9 7900X3D, RTX 4070 Super, Ubuntu 20.04, CUDA 12.4, TensorRT 10.9.0.34, OpenCV 4.10.0, BATCH_SIZE=5 Preprocess below 1 ms, inference about 2 ms, postprocess below 1 ms over 1000 samples

Citation

If you use this repository in academic work, use the GitHub "Cite this repository" entry.