ncnn_106landmarks
1. add landmark
(1). the original model comes from repositories:
https://github.com/zuoqing1988/ZQCNN by zuoqing
(2). the process of converting mxnet model to ncnn formate is:
zuoqing1988/ZQCNN#50
(3). new landmark model: landmark_big.param && landmark_big.bin has also uploaded, input size: 48x48
you can change the model name in the test code to test the model:
change "landmark.param", "landmark.bin" to "landmark_big.param" && "landmark_big.bin"
(4). add input size: 96x96 model:
you need to change the input size 48 x 48 to 96 x 96:
ncnn::Mat ncnn_in = ncnn::Mat::from_pixels_resize(img_src.data,
ncnn::Mat::PIXEL_BGR, img_src.cols, img_src.rows, 48, 48);
To:
ncnn::Mat ncnn_in = ncnn::Mat::from_pixels_resize(img_src.data,
ncnn::Mat::PIXEL_BGR, img_src.cols, img_src.rows, 96, 96);
(5).add input size: 112x112 model:
the most effect model.
2. add retinaface detection
the code refer to the repositories:
https://github.com/Charrin/RetinaFace-Cpp by Charrin
model comes from:
https://github.com/Charrin/RetinaFace-Cpp/tree/master/Demo/ncnn/models
result:
3.add jetson nano project based on vulkan
(1). build vulkan ncnn:
https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-vulkan
(2). build the project:
>mkdir build && cd build && cmake .. && make -j3
>./main
4.add mobilefacenet
5.use openmp to optimize for loops
test result:
do not use vulkan:
' start face detect. 4 faces detected. start keypoints extract. keypoints extract end. start keypoints extract. keypoints extract end. start keypoints extract. keypoints extract end. start keypoints extract. keypoints extract end. time cost: 137.495ms '
use vulkan:
' [0 NVIDIA Tegra X1 (nvgpu)] queueC=0[16] queueT=0[16] memU=2 memDL=2 memHV=2 [0 NVIDIA Tegra X1 (nvgpu)] fp16p=1 fp16s=1 fp16a=0 int8s=1 int8a=0 start face detect. 4 faces detected. start keypoints extract. keypoints extract end. start keypoints extract. keypoints extract end. start keypoints extract. keypoints extract end. start keypoints extract. keypoints extract end. time cost: 553.328ms '