Training model without landmarks

[vovaxnz]

Hi, Thanks a lot for your great work! I ran into a problem when I fine-tune the model on my own dataset with faces marked without landmarks. Can you help me please and show what I'm doing wrong?

I organized annotations in the label.txt file as follows:

# vlcsnap-2019-12-21-16h31m20s653.jpg
729 81 115 104 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 1.0
1117 549 96 101 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 1.0
273 3 71 56 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 1.0
# vlcsnap-2019-12-21-16h31m22s880.jpg
1342 347 79 64 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 1.0
625 289 128 99 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 -1.0 1.0
...

in data/config.py i changed batch_size from 24 to 6 and launched training using resnet50 during training the loss decreased, but when I made predictions on previously unseen data using python test_widerface.py predicted boxes were worse than with no-finetuned model this is how it looks before fine-tuning: and this is how they looked after fine-tuning: Could you please say what this may be connected with?

[alicera]

@moto55
hi, do you know widerface annotations format? I only know the face box and face landmark. I don't know the 0 and 0.82 Example: 0--Parade/0_Parade_marchingband_1_849.jpg 449 330 122 149 488.906 373.643 0.0 542.089 376.442 0.0 515.031 412.83 0.0 485.174 425.893 0.0 538.357 431.491 0.0 0.82

[DuckJ]

@moto55 have you solved your problem? how to train dataset without landmark

[vovaxnz]

@DuckJ I could not find a way to train without landmarks. Just used detector as is, it already works well

[DuckJ]

if the training label is organized in the above form,loss will not calculate landmark loss, right? You mean that the effect will become worse after finetune, right?

[FlyEgle]

i think the predict bbox is reverse the width and height

[watertianyi]

@vovaxnz Hello, can I ask you how to use your own data set to generate a data set in a format like widerFace. It is very urgent. It seems that no one maintains this code?

[alicera]

I think we have to know the widerface annotations format Do you know how to read it ? I don't know the 0 and 0.82 Example: 0--Parade/0_Parade_marchingband_1_849.jpg 449 330 122 149 488.906 373.643 0.0 542.089 376.442 0.0 515.031 412.83 0.0 485.174 425.893 0.0 538.357 431.491 0.0 0.82

[watertianyi]

@alicera 0 is equivalent to comma separated key points, 0.82 is estimated to be less than necessary

[alicera]

What is your label tool？

[watertianyi]

@alicera labelme