zhouyuchong / face-recognition-deepstream

Deepstream app use retinaface and arcface for face recognition.
MIT License
58 stars 12 forks source link

hi, I unable add your face landmarks(face-detect-deepstream) of you to this pipeline. Please, help me! #16

Closed manhdoan291 closed 1 year ago

manhdoan291 commented 2 years ago

I tried to include the face landmarks that I asked you @zhouyuchong before into this pipeline but it seems to have a data flow error that I can't fix at all. I don't seem to see result_landmark return results when I add queue4.link(tgie) tgie.link(queue5) queue5.link(tiler) tiler.link(queue6) main.py.txt it even lost bounding box object when i set process-mode=2(get-current). How to fix these error ? Screenshot from 2022-10-20 15-53-16

manhdoan291 commented 2 years ago

my video when run: https://drive.google.com/file/d/1hzRvY99JFt6LxU3itwjMp0DetsjCMzuA/view?usp=sharing

zhouyuchong commented 2 years ago

@manhdoan291 It seems you are using codes in another repo. I've deleted YOLOv5 from current pipeline since there are some problems in retrieving data in process-mode=2. However, if you don't need to extract face embedding but only to visualize detection and landmarks, you can refer to this repo which I test and modify today.

manhdoan291 commented 2 years ago

@zhouyuchong I will remove yolov5 from this pipeline. I'm trying to combine (face landmark -> face-align -> embedding) but can't because of wrong pipeline. I don't know how to fix it ? I don't seem to see result_landmark return results(vscode of me) on this repo.

zhouyuchong commented 2 years ago

@manhdoan291 In this repo, you can't reach lmks in python codes because I retrieve these data in c++, gst-nvinfer source codes. Deepstream python only offers few APIs which can't satisfy alignment requirement.

manhdoan291 commented 2 years ago

@zhouyuchong
so is there any solution to approach lmks in this python code when retrieving these data in c++ source code, gst-nvinfer for embed faces?

zhouyuchong commented 2 years ago

@zhouyuchong so is there any solution to approach lmks in this python code when retrieving these data in c++ source code, gst-nvinfer for embed faces?

What's your purpose to get landmarks in this repo? Since I use fakesink and drop faces once it's been infered by arcface.

For you question, I think you can get those data the same way as before. Using a probe in python codes to get tensor-output.