search

antoalli / Deco

DeCo: Denoise and Contrast for Category Agnostic Shape Completion
34 stars 3 forks source link

Question About the paper #8

Closed zztzzt200 closed 2 years ago

zztzzt200 commented 2 years ago

Hi ! I don't quite understand what 'frame' stands for. Could you please tell me something about its function. Thanks a lot

antoalli commented 2 years ago

Hi! With the word "frame" we intend the portion of the input partial point cloud which is around the missing part. If you take a look at the right part of the teaser (paper Fig. 1 or the image in this repo) the frame points are the yellow ones. If you take a look at our architecture in Fig. 2 of the paper you will notice that the decoder works in a hierarchical way: it first predicts a coarse output resembling the frame + the missing part points (Y_fm) and then specializes the prediction to the only missing part (Y_m). Obviously, this is a regularization strategy we can employ during training for better merging the predicted missing part with the partial input, at test time we simply disregard Y_fm and consider only the missing part prediction Y_m.

See also our paper for a detailed explanation and feel free to ask for more info if something is still unclear, Antonio

zztzzt200 commented 2 years ago

Oh! I get it. Figure 1 made me think that the frame was part of the input, which conflicts with the following text and confused me. I understand now. Thank you very much.