Open ashuezy opened 6 years ago
Just create by using https://github.com/AlexeyAB/Yolo_mark or find dataset http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/ or http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html with labeled faces and convert labels to Yolo v3 format. Then train Yolov3-Tiny or Yolov3: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects
This is enough for counting faces on the images:
To process a list of images data/train.txt and save results of detection to result.txt use:
darknet.exe detector test data/voc.data yolo-voc.cfg yolo-voc.weights -dont_show < data/train.txt > result.txt
I ran Yolov3-tiny and Yolov3. But both of them miss out on medium to long range people. This is for the crowd counting application. I want to increase accuracy and detect far off people too. Need accurate counting of people with ±90 accuracy. I am a little confused with the logic. I think i will have re annotate datasets with face-detection/head detection. Still not sure how to implement this. Kindly advice with relation to accuracy.
Thank you.
Sample Images to be targeted:
https://s15130.pcdn.co/wp-content/uploads/2018/01/2018-01-20T180225Z_1_LYNXMPEE0J0OM_RTROPTP_4_USA-TRUMP-WOMEN.jpg
https://i.ndtvimg.com/i/2018-04/asaram-bapu-delhi-ashram-650_650x400_41524415410.jpg
https://scontent.fdel8-1.fna.fbcdn.net/v/t1.0-9/28056662_10155601483857669_5886305188712316563_n.jpg?_nc_cat=0&oh=c5b7db655e221c5ea58ca6b94be89b58&oe=5B938D02
https://scontent.fdel8-1.fna.fbcdn.net/v/t1.0-9/28056662_10155601483857669_5886305188712316563_n.jpg?_nc_cat=0&oh=c5b7db655e221c5ea58ca6b94be89b58&oe=5B938D02