Geometry and texture

The geometry and texture of the each model is available in original Collada format and in converted OBJ/MTL format. The converted OBJ have been aligned and size normalized. For ShapeNetCore v1, the OBJ/MTL files were converted using a Blender-based pipeline and there are known issues with missing normals, right-left flipping, and (what else?) that can cause problems with rendering. For ShapeNetCore v2, the OBJ/MTL files were converted using the Open Assimp Import Library, fixing many of the above issues. Original mesh and material structure of each model can be viewed using the parts panel in the model viewer.


In addition, we provide precomputed solid and surface voxelizations of each model. Resolution is 128. Voxelizations are computed using binvox. To view the voxelization, open the parts panel in the model viewer and select voxels-solid or voxels-surface.


For each model, we generate screenshots of the model from the 6 canonical viewpoints and 8 turn table images. These screenshots can be viewed using the images panel in the model viewer.


Models in ShapeNet are linked to ImageNet and WordNet. WordNet is a widely-used, English lexical database that groups words into cognitive synonyms (synsets). WordNet synsets are interlinked with various relations, such as hyper and hyponym, and part-whole relations. We identify the WordNet synset by the WordNet offset (0 padded to 8 digits), we refer to this as the wnId. Note that the offsets are for WordNet 3.0. We provide a mapping of WordNet 3.0 synset to WordNet 3.1 synsets. Use the taxonomy view to browse ShapeNet using the WordNet taxonomy.

The following fields are for categories

  • wnhypersynsets all wnIds (including hypernyms) associated with the model
  • wnhyperlemmas words corresponding to the wnhypersynsets
  • wnsynset wnIds directly associated with the model
  • wnlemmas words corresponding to the wnsynset

For ShapeNetSem, we also label models using a smaller manual category taxonomy. This taxonomy is simpler than WordNet, covering mainly furniture and electronics (and includes finer grain classification of electronics lacking in WordNet).


We try to provide consistent up/front alignment for all models in ShapeNetCore and ShapeNetSem.

The following fields indicates the up/front axis for the KMZ models

  • up vector indicating direction of up axis
  • front vector indicating direction of front axis


We provide real world size estimates for ShapeNetSem models. Real world sizes are useful for using models to construct scenes.

The following fields are for sizes:

  • scale unit the original Collada model was likely modeled in
  • aligned.bbdims vector of estimated bbdims of aligned model