Source of testing images in the paper about zero123

[ruoshiliu]

Hi @XinyangHan , I took a bunch of pictures of objects in our lab and my apartment as those objects are most likely not in the objaverse dataset. In addition, we find images on the internet as well. But these are mostly for qualitative experiments. For a 3D testing dataset with 3D ground truth, the available datasets are probably much more limited. To name a few that are on top of my mind, there are GSO, CO3D, RTMV, ShapeNet. In terms of choosing which objects to use within the dataset, we just randomly sample a subset and run experiments on those. We didn't cherry-pick within a dataset.

[XinyangHan]

Got it!!! Thanks so much for replying!! Btw, when testing on in-the-wild images, it seems that the segmentation would have a little artifacts like shown below. May I ask if you have some advice on matting in the wild images as input or plan to update the preprocessing module~
Thanks again for your kindly help~

[basilevh]

Hi @XinyangHan , the foreground segmentation model that we use was not created by us -- instead, we apply CarveKit off the shelf. The goal of this preprocessing step is simply to turn the background white. If the quality is not good enough, feel free to segment the object manually, or perhaps try using another model such as Meta's Segment Anything instead. In that case you should turn off the "preprocess image" checkbox in our Gradio demo, and center the foreground object against a white background after segmenting itout. Hope this helps!

[XinyangHan]

Thanks so much for your help!!!