DeepOcclusion icon indicating copy to clipboard operation
DeepOcclusion copied to clipboard

Could you give some suggestions about how to train on my own dataset

Open scut-salmon opened this issue 7 years ago • 5 comments

How can I obtain my own pickle files in VGG/models base on my dataset ?

scut-salmon avatar Feb 11 '18 16:02 scut-salmon

Hi, I pushed some files to train new models on different dataset. You can take a look at the guide and try the files mentioned in it. Let me know if you have any error or problem when trying it.

gooa1121 avatar Feb 12 '18 16:02 gooa1121

@gooa1121 Thank you for your work. When retraining the Gaussian and regression model, what labels do I need to use? The body parts mask? like background subtraction retraining?

lc-git avatar Mar 21 '18 07:03 lc-git

Following the section 3.2.2 in the main paper, the label is the normalized displacement between the coordinate of the pixel and the center of the bounding box it belongs to. For the pixels not belonging to any bounding box, you can give it any number <= -1

gooa1121 avatar Mar 21 '18 14:03 gooa1121

As you mentioned, label is the displacement between the pixel and the center so I think label has 2 channels. Why the shape of parts labels is (H, W, 5)? What does "5" mean? Thank you.

lc-git avatar May 14 '18 12:05 lc-git

The first channel indicates foreground or background, 2nd and 3rd are the displacement to the center, 4th and 5th are the displacement to the right, bottom corner of the bounding box.

gooa1121 avatar May 15 '18 19:05 gooa1121