pcb9382 / facealgorithm Goto Github PK

hi, I have some questions about generate the coordinates of keypoints. Why Face_Alignment_prob need to plus1 and then divide 2? What is the meaning of Face_Alignment_prob

alignmentface.landmarks[start]=int((Face_Alignment_prob[start]+1)*img.cols / 2); alignmentface.landmarks[start+1]=int((Face_Alignment_prob[start+1]+1)*img.cols / 2);

Thanks for your help.

When I tried to convert my onnx file to trt file (tensorrt7), I got the error below, how can I solve it?

[12/11/2023-16:10:27] [I] [TRT] ModelImporter.cpp:135: No importer registered for op: ScatterND. Attempting to import as plugin.
[12/11/2023-16:10:27] [I] [TRT] builtin_op_importers.cpp:3659: Searching for plugin: ScatterND, plugin_version: 1, plugin_namespace:
[12/11/2023-16:10:27] [E] [TRT] INVALID_ARGUMENT: getPluginCreator could not find plugin ScatterND version 1
ERROR: builtin_op_importers.cpp:3661 In function importFallbackPluginImporter:
[8] Assertion failed: creator && "Plugin not found, are the plugin name, version, and namespace correct?"
[12/11/2023-16:10:27] [E] Failed to parse onnx file
[12/11/2023-16:10:27] [E] Parsing model failed
[12/11/2023-16:10:27] [E] Engine creation failed
[12/11/2023-16:10:27] [E] Engine set up failed

网盘链接失效了，被百度视为违规内容

When I provide images in a loop without any delay, the processing time for yolov7-face or yolov8-face is short. However, when I feed the images to the detection function one by one, introducing a 1-second time interval between each photo, the processing time becomes longer. What might be causing this issue?

Here are the processing times for images in a loop:

../images//test1.jpg average: 389.05ms
../images//test3.jpg average: 134.054ms
../images//cam4.jpg average: 104.824ms
../images//test11.jpg average: 93.1855ms
../images//test7.jpg average: 86.4966ms
../images//test8.jpg average: 85.9823ms
../images//arac2.jpg average: 67.5789ms
../images//arac3.jpg average: 69.3688ms
../images//arac4.jpg average: 68.7759ms
../images//test9.jpg average: 75.8391ms

And here are the processing times with 1-second intervals between images:

../images//test1.jpg average: 267.529ms
../images//test3.jpg average: 313.996ms
../images//cam4.jpg average: 159.6ms
../images//test11.jpg average: 315.25ms
../images//test7.jpg average: 296.985ms
../images//test8.jpg average: 237.869ms
../images//arac2.jpg average: 206.976ms
../images//arac3.jpg average: 244.924ms
../images//arac4.jpg average: 185.883ms
../images//test9.jpg average: 239.323ms

Upon analyzing the detect function, I've identified that the following line is taking a long time:
CHECK(cudaMemcpyAsync(decode_ptr_host[i],decode_ptr_device,sizeof(float)(1+MAX_OBJECTSNUM_BOX_ELEMENT),cudaMemcpyDeviceToHost,stream));

What could be the issue and what can be the solution? CudaMemCpy is slower when images are given one by one. How can I solve this?

如题。

The Baidu link isn't working, can you share the models via other services like Google Drive or One Drive?

你好，我想问下，计算人脸角度时，是不是可用关键点数量越多越好