// the find
jkjung-avt/tensorrt_demos
TensorRT MODNet, YOLOv4, YOLOv3, SSD, MTCNN, and GoogLeNet
A set of TensorRT inference demos for NVIDIA Jetson boards (Nano, TX2, Xavier NX) and some x86 GPUs, covering GoogLeNet, MTCNN, SSD, YOLOv3/v4, and MODNet video matting. It is a working reference for getting Caffe, UFF, and ONNX models into TensorRT engines on older JetPack releases, not a maintained inference library for current TensorRT.
- The YOLOv3/v4 section is the most useful part. It gives mAP and FPS on Nano for each variant, plus INT8 and DLA numbers on Xavier NX, so you can pick a model size before spending half an hour building an engine.
- The INT8 section reports its failures: yolov4-608 INT8 drops from 0.488 to 0.317 mAP, and DLA engines for yolov4-tiny can't be built. Most demo repos leave that out.
- The yolo_layer plugin is a real C++ TensorRT plugin built on IPluginV2IOExt, with a Makefile and Cython bindings in pytrt.pyx. It shows how a custom layer gets wired in, instead of relying only on the stock ONNX parser.
- Tested only against TensorRT 5 through 7.x and JetPack 4.x. The SSD path depends on UFF, which TensorRT 10 removed, and the README says nothing about newer TensorRT versions. Expect to patch the code before it runs on a current JetPack.
- The dependency setup is the weakest part. Demo #3 needs TensorFlow 1.x with a UFF converter pinned to tensorflow 1.12. Demos #4 and #5 pin onnx==1.9.0 and recommend compiling protobuf 3.8 over a couple of hours. Each model directory has its own install_pycuda.sh (modnet, ssd, yolo), so getting one demo running means reading that demo's section first.
- The licensing is mixed. The bundled MODNet weights are CC BY-NC-SA 4.0, so they are off-limits for commercial use, and the MTCNN code has no license specified. The MIT license on the rest of the repo does not cover either of them, and the license section is the only place this is spelled out.