Whether this dataset can be used for NCOS-based method?

[hzqtitanium]

Dear Authors, This dataset is a great one in the field of category-level pose estimation. And I would like to know that would you provide a version containing NOCS map so researchers can try the NOCS-based method on it? Best Ziggy

[swtyree]

Hi @hzqtitanium, thanks for reaching out. We haven't discussed creating NOCS maps, but we'd be happy to have further discussions. Within each category, the reference mesh for each instance should be in a canonical position+rotation for that category, which should make it straightforward to normalize. Are you aware of any scripts for rendering the NOCS map for a given GT mesh + pose? I found a few relevant posts on the NOCS repo, including this one, but I likely won't have time to try to code at the moment.

[taeyeopl]

Does the same category align with canonical rotation (e.g., mug1 and mug2 have the same rotation in handle parts)? Or does each instance have a unique rotation, regardless of the category relation?

[swtyree]

(Apologies for a first version of this reply that was geared toward the HOPE dataset instead of HANDAL. I had just checked some issues on that repo and forgot which repo I was on. 😑) The objects should be reasonably well aligned by category in HANDAL. The alignment was done by hand and we didn't have big discussions about it. But the intention was they they should be aligned logically by category.

[taeyeopl]

Thanks for the quick replies 😄