Landmarks predict? about deepfashion2 HOT 5 OPEN

A keypoint’s location is modeled as a one-hot mask. In this way, you can easily train landmark estimation network with your framework. Instead of predicting K masks for K categories given a detected bouding box, you need to predict K masks, one for each of K keypoint types.

from deepfashion2.

@geyuying
Thanks for the suggestion

I have already trained landmark keypoints with mtcnn main.py
But I would like to know how to visualize landmarks

from deepfashion2.

https://github.com/facebookresearch/Detectron/blob/master/detectron/utils/vis.py def(vis one image)

from deepfashion2.

@geyuying
Thanks for the suggestion

I have already trained landmark keypoints with mtcnn main.py
But I would like to know how to visualize landmarks

can you share your code about landmark keypoints? thanks.

from deepfashion2.

A keypoint’s location is modeled as a one-hot mask. In this way, you can easily train landmark estimation network with your framework. Instead of predicting K masks for K categories given a detected bouding box, you need to predict K masks, one for each of K keypoint types.

hi, @geyuying @TDBTECHNO ,
Does K equal to 294? Do I have to convert keypoints of every cloth type into 294 heatmaps as supervision? I tried to train the simple Hourglass module based network by using the converted 294 heatmaps but the network can't convergence. And I also tried OHKM.
It seems that the deepfashion2_to_coco.py concatenates all 294 keypoints together for each cloth. That is, the maximum number of valid keypoints is only 39(long sleeve outwear). How to handle this unbalance between negative and positive labels?

from deepfashion2.