Datasets:
Missing symmetry annotations in models_info.json
The models_info.json file is present, but it doesn't contain the symmetry fields for objects 1, 3, and 5 — the same three objects that are marked as symmetric in the i6DL-Edge code. These fields are required by the BOP evaluation pipeline.
For objects 1 and 3 the symmetries look straightforward, but I'm struggling with object 5.
Could you share the symmetry annotations (or a complete models_info.json)? Any help would be much appreciated.
Thanks in advance!
Hi, thanks for reaching out and for your interest in the IndustryShapes dataset.
Objects 1 and 3 indeed have reflection symmetries. However, regarding the BOP definition of discrete symmetries, these objects don't cause 6D pose ambiguities (i.e., you can't rigidly rotate them into indistinguishable poses). This applies for all the objects. It is worth noting, though, that depending on observing distances and lighting conditions, visual ambiguities may still arise due to a lack of texture (e.f. obj 4).
The correct way to evaluate them is as non-symmetric. Regarding the BOP evaluation pipeline, it simply calculates the appropriate metrics depending on the symmetries specified for each object. If an object is treated as non-symmetric, the toolkit's symmetry-aware metric become their standard, strict counterparts. Under the hood, it evaluates against the single valid ground truth pose—meaning MSSD becomes exactly the same as standard MSD, and the same applies for MSPD.
Furthermore, even if you include these objects under symmetric_obj_ids in your dataset_params.py, the toolkit will look for the symmetry transformations using the get_symmetry_transformations function. If it doesn't find any in the annotations, it defaults to the identity transformation.
Let me know if the evaluation pipeline is throwing a specific error on your end and we can look into it.
Thanks again!