ccvl/3D-DST-models

3D-DST-models

As part of our data release in 3D-DST, we present aligned CAD models for all 1000 classes in ImageNet-1k.

See wufeim/DST3D for synthetic data generation with 3D annotations using the CAD models here.

Besides the .csv file as visualized in the dataset viewer above, we also provide a python script (models_3d_dst.py) to help integrate with other Python modules.

Fields

For each CAD model, there are seven fields:

• synset: synset associated with each ImageNet-1k class.
• model: model ID from ShapeNet, Objaverse, or OmniObject3D
• distance: corrected distance -- our rendering code normalize object sizes automatically, correct the rendering distance if the normalization fails
• azimuth and elevation: viewpoint parameters used to align the canonical pose of all CAD models
• strength: 1 if the current CAD model is good; 2 if the current CAD model only approximates the current synset
• sampling: one of the four sampling strategies here:
  • 1: viewing from all azimuth angles, and all elevation angles, e.g., airliner
  • 2: viewing from front azimuth angles, and all elevation angles, e.g., microwave, washing machine
  • 3: viewing from all azimuth angles, and only elevation angles when viewing from up to down, e.g., keyboard, bath tub, car, beach wagon
  • 4: viewing from front azimuth angles, and only elevation angles when viewing from up to down, e.g., printer, stove