I have a question about getting the outputs of Detectron2 Denspose.

[frankkim1108]

I am trying to use the ApplyNet and get the estimated U,V coordinates. I successfully did apply_net show and got the result visually. However I want to get the coordinates. I tried using apply_net dump and got a results.pkl file. However when I tried to open the pkl file from Pycharm I got an error that says "ModuleNotFoundError: No module named 'densepose.structures'" It works fine when I open the file with terminal.

I want to know where I can get the coordinates from a file.

[ppwwyyxx]

If it works in terminal then it's a limitation of Pycharm which we'll not investigate.

[frankkim1108]

I apologize for my question being not clear. My question is about where can I find 3D coordinates when I do python apply_net.py dump configs/densepose_rcnn_R_50_FPN_s1x.yaml DensePose_ResNet50_FPN_s1x-e2e.pkl "Jongwook_F2.jpg" --output results.pkl -v When I opened the results.pkl file I couldn't see the coordinates. Thank you for reading my questions.

[vkhalidov]

The output file results.pkl should contain the list of results per image, for each image the result is a dictionary. DensePose results are stored under pred_densepose key, it is an instance of DensePoseResult class. It contains the list of results in the results attribute. Each result is encoded as a string. So to extract the result one needs to decode the string using DensePoseResult.decode_png_data, e.g.

(densepose_shape, densepose_data_encoded), bbox_xywh = densepose_results[i]
iuv_arr = DensePoseResult.decode_png_data(densepose_shape, densepose_data_encoded)

[frankkim1108]

Thank you for your response. However, due to my lack of knowledge I couldn't follow your steps properly.

To open the pkl file I used

with open('/home/elysium/detectron2/projects/DensePose/result3.pkl', 'rb') as f:
    data_list = []
    while True:
        try:
            data = pickle.load(f)
        except EOFError:
            break
        data_list.append(data)

1. When I first opened the results.pkl file I got. The picture had only one person in it.

   {'file_name': 'Jongwook_F2.jpg', 'scores': tensor([0.9999, 0.3534]), 'pred_boxes_XYXY': tensor([[  10.4258,  358.9859, 2114.6133, 2578.4807],
       [ 816.5173,  855.3380, 1347.3788, 2662.1370]]), 'pred_densepose': <densepose.structures.DensePoseResult object at 0x7f841b9f9150>}

2. Next I tried to get the pred_densepose by following your advice.

print(data_list[0][0]['pred_densepose']

and I got

DensePoseResult: N=2 [[3, 2219, 2104], [3, 1806, 530]]

3. This is how far I went. However, I couldn't figure out how to use DensePoseResult.decode_png_data nor

   (densepose_shape, densepose_data_encoded), bbox_xywh = densepose_results[i]
   iuv_arr = DensePoseResult.decode_png_data(densepose_shape, densepose_data_encoded)

   I'm very sorry for asking too much... I was frustrated because I got stuck on this stage for a long time.

[frankkim1108]

https://github.com/facebookresearch/detectron2/issues/165#issuecomment-561943932 I got answers from here Thank You for helping me!