Training BANMo using my own videos

[Xin-97]

Training BANMo using my own videos

[gengshan-y]

Hi, the detection pipeline in the repo is the one I used to generate the preprocessed data, besides some difference in color visualization. To get a better color vis, here is some discussion in the original densepose repo. I can look further if you post the bad results.

[Xin-97]

Hi, as shown in the following, I preprocess cat-pikachiu with the raw data (the second one), the bounding box is always not correct and the color does not look like the preprocessed data. What does the different color mean in the shown video (the first video)?

vis.mp4

vis_1.mp4

[gengshan-y]

The bbox on the cat is correct as far as I can see. The bbox on other categories are automatically filtered and not used for reconstruction.

The vis in the proposed data is a densepose visualization that I posted in the first reply.

Let me know if this does not answer your question.

[Xin-97]

Hi, as shown in the second video, the bounding box always has a bit more on the bottom and right. This happens in all videos.

[gengshan-y]

You are right, this is a separate visualization bug. This line should be replace by

    outputs.pred_boxes.tensor[:,:2] -= pad
    outputs.pred_boxes.tensor[:,2:] -= 2*pad

I don't think this affects reconstruction though. Thanks for the catch. Will update the repo.

[Xin-97]

Hi, I have tried the code, the second line may also be -= pad. Thank you very much!