Open faderani opened 3 years ago
Manideep08
commented
3 years ago Could you please elaborate a bit more, for instance, what are the shapes of hard case faces in your dataset(just the bounding boxes), and what are the different param_bbox_list values you have tried?
Did you try changing the feature_map_size list and training it?
Input shape of your image and if possible share your configuration file so that I will try to help you.
faderani
commented
3 years ago Most of the faces in my dataset are no larger than (30,30)... the quality of the image itself is not good too so lot of hard cases.
I tried changing param_bbox_small_list = [10, 20, 40, 80, 160] to param_bbox_small_list = [10, 20, 30, 40, 50]
and the upper bound from param_bbox_large_list = [20, 40, 80, 160, 320] to param_bbox_large_list = [20, 30, 40, 50, 60]
input shape of my image is either 360x360 or 320x240 or 240x320. I didn't change any other thing.
and I don't want to train it on my dataset. I just want to train it on WIDERFACE in a way which can get more robust to tiny faces.
licc0431
commented
3 years ago What I detect is the face in the picture of about 100*200,and most of the faces in my dataset are no larger than (30,30).Did you find a good solution in the end, or use other model training?
Hi, How should I change the architecture of network to make it robust to hard cases (small faces). In my scenario faces are mostly tiny. I reduced param_bbox_small_list and param_bbox_large_list and retrained the network but I'm getting no detections at all.