YOLOv3-TensorRT-INT8-KCF is a TensorRT Int8-Quantization implementation of YOLOv3 (and tiny) on NVIDIA Jetson Xavier NX Board. The dataset we provide is a red ball. So we also use this to drive a car to catch the red ball, along with KCF, a traditional Object Tracking method.

• See yolov3/requirements.txt.

• TensorRT >= 7.0.0 (Pre-installed on NX.)
• OpenCV and opencv_contrib == 3.4.0 (See here for installation help.)

On GPU server, not NX.

git clone https://github.com/lingffff/YOLOv3-TensorRT-INT8-KCF.git
cd YOLOv3-TensorRT-INT8-KCF
cd yolov3
# Download official pre-trained COCO darknet weights
sh weights/download_yolov3_weights.sh

Download redball dataset here, unzip and replace the folder redball. Then start training. Remove '--tiny' if you train YOLOv3 model.

python train.py --device 0 --tiny

Now we get YOLOv3(tiny) weights in weights/best.pt. Transfer it to binary file redball(-tiny).wts, which convert weights to TensorRT for building inference engine.

python gen_wts.py --tiny

Then copy ./redball(-tiny).wts to NX Board.

On NX Platform below.

git https://github.com/lingffff/YOLOv3-TensorRT-INT8-KCF.git
cd YOLOv3-TensorRT-INT8-KCF
# Put redball(-tiny).wts in YOLOv3-TensorRT-INT8-KCF

Build the project.

mkdir build
cd build
# YOLOv3: -DTINY=OFF, tiny: -DTINY=ON
cmake -DTINY=ON ..
make -j$(nproc)

Now we get executable file build_engine and detect.
Run build_engine. Use -s argument to specify quantization options: int8, fp16, fp32(default).

./build_engine -s int8

Run detect to detect pictures or camera video. You can also check KCF tracking method here by other options below.

./detect -d ../samples

Options:

Wow! FP16 is amazing!!!

• Convert weights to TensorRT by a more common way, like ONNX.
• Run detection and tracking multi-thread-ly.
• Implement a Quantization & Inference framework myself.

YOLOv3 Pytorch implementation from ultralytics/yolov3.
YOLOv3 TensorRT implementation from wang-xinyu/tensorrtx.
TensorRT Int8 implementation from NVIDIA/TensorRT/samples/sampleINT8.

With my sincerely appreciation!

Just call me Al (not ai but al. LOL.) / Albert / lingff.
E-mail: [email protected]
Gitee: https://gitee.com/lingff
CSDN: https://blog.csdn.net/weixin_43214408