Hi nihui,

I see that in generate_anchors, the largest anchor is 512*512 (base size=256, scale=2.f).

I want to change the size of the anchor to detect larger faces.

But I can't detect a larger face by increasing base_size or scale.

What should I do to detect larger face sizes?

BR

@nihui 大佬好，感谢大佬的代码，我已经按照设置把ncnn_dir 和 opencv_dir设置为自己的了，并且所有的ZIP下载下来解压后都放在了jni下，我的cmakelists像这样：

project(scrfdncnn)

cmake_minimum_required(VERSION 3.10)

set(OpenCV_DIR ./sdk/native/jni)
find_package(OpenCV REQUIRED core imgproc)

set(ncnn_DIR ./arm64-v8a/lib/cmake/ncnn)
find_package(ncnn REQUIRED)

add_library(scrfdncnn SHARED scrfdncnn.cpp scrfd.cpp ndkcamera.cpp)

target_link_libraries(scrfdncnn ncnn ${OpenCV_LIBS} camera2ndk mediandk)

但是在编译项目的时候还是找不到ncnn,想问下错在哪里了～感激不尽

请问模型要怎么手动修改yolov5s 模型，有推荐的工具或者需要遵循的规则吗？或者可以出一个文档吗

Hi,

would you mind to pack a linux (intel/arm/gpu) version of the scrfd ?

would love to see arm / linux perfoamance

Best

The problem

Actually this is not a problem with your code or ncnn model it just a generic problem when it comes to face heavy operations. Slowness and image frame cuttering because you detect faces in every camera frame

So

Well actually phones cpus and gpus aren't much power than our desktops or laptops. It needs a different treatment and workaround to get the best results so far I have used many workarounds and really get some good results overall but still need much smoothness between frames, So far I tried

• Resizing an image before getting into processing
• Do the process task in a standalone thread
• Trackers ?
• If you know any workaround please mention it

How to get more good results

I believe using trackers play a powerful role in this problem but in myself didn't find a good resource to start implement that and see what the output. Did you ever tried to solve this problem using trackers? and if you have a good repos for trackers to use please mention them below

Thanks!

Hi

When we try to install apk it gives INSTALL FAILED OLDER SDK. error.

undefined reference to `cv::imwrite(std::__ndk1::basic_string<char, std::__ndk1::char_traits, std::__ndk1::allocator > const&, cv::_InputArray const&, std::__ndk1::vector<int, std::__ndk1::allocator > const&)'
clang++: error: linker command failed with exit code 1 (use -v to see invocation)

I am using your opencvmobile sdk.

How we can compile for android API 22 ?
Due to our chipset we stucked API 22

Best

用那个benchmark只用大核帧数不如全核，这个scrfd就有问题

Android studio4.2.1
ncnn-20210525-android-vulkan
opencv-mobile-4.5.1-android