possible to train without mask?

[cchamber]

I have a new dataset I would like to train on but do not have segmentation masks. Is it possible to train without the masks?

[abinjoabraham]

Same question I am also having. Can we train without the masks. Actually what is the main need for the masks ? Can we implement the 2D pose by just the key points?

[cchamber]

I trained using a bounding box instead of a mask and gave the coordinates of the box instead of the segmentation polygon. I only used images without occlusion for training. This seems to work.

[abinjoabraham]

@cchamber thanks buddy. Can I look on your code ? Is the fork available online. ?

[murrayLuke]

@abinjoabraham the masks seem to be used to filter out unlabeled people from the dataset. For example if there is a crowd of people maybe three of them will have body parts labeled. The mask then covers the remaining people in the crowd. When the algorithm recognizes a body part on someone who is masked out, it is ignored, since there is no ground truth.

[typhon3d]

Thus, when the dataset includes only images with one single person, the mask is redundant? just a full white image can be used for the mask and the results of the model will be of equal accuracy?

[cchamber]

@abinjoabraham I made a couple of edits to Anatolix's code https://github.com/cchamber/openpose_keras
There's also an annotation tool for video. you can convert your labels made with that to COCO for training
https://github.com/cchamber/convert_videolabels_to_NNinput
if you want to build a mask you can take the area included within the keypoints (convex hull) with a bit of padding around it

[abinjoabraham]

@cchamber Thank you for sharing your version. I will surely study on it. Being a newbie to python and deep learning world these codes are giving me very good knowledge in Artificial Intelligence.

[1209805200]

@cchamber the mask is necessary? i am not sure the groundtrue include what?keypoint and mask??? hope your reply