search

yizhiwang96 / Slice3D

Slice3D: Multi-Slice, Occlusion-Revealing, Single View 3D Reconstruction
MIT License
91 stars 8 forks source link

Different results for demo picture #2

Open SheenaB2 opened 3 months ago

SheenaB2 commented 3 months ago

Hi,

Thank you for your excellent work on the Slice3D project! I am trying to run the demo picture of a shoe using the code provided with your pre-trained models. I followed the instructions in the "Setup" and "Testing on Single Image" sections of your README file. However, I obtained different results compared to what is shown in your paper using both the regression-based method and the generation-based method. Specifically, the shape appears different, and for the regression model, there are no holes inside the shoe. I attached screenshot of my results below.

Could you please let me know if there is anything I might be doing wrong? Thank you! regression_based_shoe generation_based_shoe

yizhiwang96 commented 3 months ago

Hi @SheenaB2, thank you for your interest in Slice3D! Your produced results are correct.

About your question: why are the results different to the ones in paper?

Ans: I did some data cleaning when releasing the code and ckpts, making the training dataset smaller to the one used in our paper. Less examples of shoes/boots could be the reason why our method failed in reconstructing the hole in regression-mode.

I also notice that our regression-based results are generally better than generation-based ones. The main advantage of generation-based slicing is that it can produce multiple results instead of deterministic ones.

About data cleaning: when slice a 3D model, we need the meshes in this 3D model to be merged. However, Blender could fail in merging meshes for some 3D models from Objaverse, yielding broken structures. I did not include these 3D models in the training dataset.