2021.06.06 Pytorch version: https://github.com/HansRen1024/C-OF. Code is released (20210816).
Some guys are not familiar with ncnn, and compile this repo with an error of could not find net.h. Please move to https://github.com/Tencent/ncnn to find what is ncnn and how to install it. Download the version we used from https://github.com/Tencent/ncnn/archive/refs/tags/20180830.zip.
I extremely optimized the code, all useless contents were removed. Right now, tracking speed is approximate 3ms. Besides, I deleted initialization funtion and OpenCV 3.x is supported now.
RK3399 20+ ms/frame
I optimized OpenTLD making it run faster and better for face tracking.
This version of TLD is faster and more stable than that in OpenCV. I delete some funtions to make it run faster. What is more, use RNet to judge the face that TLD produced to avoid TLD tracking a wrong target. In order to get a stable bounding box, I fix the width and height that MTCNN provides. Running time on my PC(Intel® Xeon(R) CPU E5-2673 v3 @ 2.40GHz × 48) is about 16ms(MTCNN, ncnn), 30ms(TLD initialization), 10ms(TLD tracking) on an image of 320*240 resolution. Besides, MTCNN can be replaced by PCN or any other face/object detection algorithms.
中文介绍地址:https://blog.csdn.net/renhanchi/article/details/85089265
OpenCV 2.4.X is required!(Now OpenCV 3.x is supported)
Install ncnn firstly, and reset ncnn's include and lib pathes in CMakeLists.txt.
mkdir build
cd build
cmake ..
make
cd ..
./demo
https://github.com/Tencent/ncnn