search

GAP-LAB-CUHK-SZ / deepFashion3D

Deep Fashion3D: A Dataset and Benchmark for 3D Garment Reconstruction from Single Images (ECCV2020)
https://kv2000.github.io/2023/06/20/deepFashion3DV2/
Other
183 stars 15 forks source link

How does GCN handle different set of input feature lines #1

Open longbowzhang opened 3 years ago

longbowzhang commented 3 years ago

Hi @kv2000 , Superb work and dataset, which in my opinion is a big contribution to this research area. Meanwhile, I have several questions. [1] some inconsistency As mentioned in the abstract, there are 563 garment instances. But according to table 1, the total number is 599. I suppose you remove some garment instances. But why, e.g., poor SMPL fitting results?

How many random views are used for rendering synthetic images, 5 (section 4.1) or 3 (section 4.3)? I admit that this does not really matter but somewhat confusing.

[2] Why ignore the shape (i.e., beta) parameter of the SMPL model from the whole pipeline? The scale of a kid’s cloth is different from that of an adult’s cloth.

[3] The gcn part. In the original Pixel2Mesh paper, the input is always the fixed ellipsoid from which the deformations are. But in your paper, the input is the varying set of feature lines depending on the category of the cloth (as shown in Fig. 3), right? How does the GCN manage to handle varying set of feature lines?

Thanks very much!

longbowzhang commented 3 years ago

Hi @kv2000 I am wondering whether it is possible that you spend some time to help me clarify the above questions?

pruepei commented 3 years ago

Hi @longbowzhang How can you access the form page at https://forms.gle/1SMrPHCyS3m7pkXM6? I failed even via v2ray ...

kv2000 commented 3 years ago

Hi @longbowzhang How can you access the form page at https://forms.gle/1SMrPHCyS3m7pkXM6? I failed even via v2ray ...

Hello longbow. Thank you for your interest to our work as well as our dataset & Sorry for keep you waiting for so long :D. [1] For the number for garments instances, we'd check out the number of the garment instances invovled and modify the description about the dataset later. As for SMPL fitting, could you provide more details about how you fit SMPL to the garment point clouds? [2] We use 5 random views rendered for our experiment, i bet you'd try to render the images in more views. [3] We ignore the shape parameters as we only adopt smpl for pose deformation to cast a initialization to our template and leave the other deformations to the following steps. [4] For the GCN part, as we mentioned in our paper, the feature lines are several disjoint closed curves predefined on the template, a new template (the feature line curves) rather than the ellipsoid should be feed to the network as input. Varing set of vertices can be handled with masks as shown in the paper.

kv2000 commented 3 years ago

Hi @longbowzhang How can you access the form page at https://forms.gle/1SMrPHCyS3m7pkXM6? I failed even via v2ray ...

Hello pruepei, the form is placed on google form; perhaps you can try whether your network status is ok by browsing google...

longbowzhang commented 3 years ago

Hi @kv2000 Thank you so much for your detailed reply. Just a minor question, as mentioned in the paper

we first determine the activated subset of feature lines according to the estimated cloth category and only feed the activated ones into the network

I suppose you feed the whole set of the feature lines but associate with a mask to denote activations, right? I am wondering [1] the mask is part of the input to the network or [2] mask is used only in the losses to make sure only the losses of the activated feature lines to backpropagate?

BTW do you have any plan to release the code? Thanks a lot.