3DCOMPAT++: An Improved Large-scale 3D Vision Dataset for Compositional Recognition
Core Concepts
3DCOMPAT++ is a multimodal 2D/3D dataset with detailed part and material annotations, facilitating compositional recognition in 3D vision tasks.
Abstract
The dataset contains 10 million stylized 3D shapes with fine-grained part and material information.
Grounded CoMPaT Recognition (GCR) task introduced to recognize shape categories and part-material pairs.
Various existing datasets compared, highlighting the unique features of 3DCOMPAT++.
Experiments conducted on shape classification, part segmentation, and GCR task performance.
Challenge organized to benchmark methods on the GCR task.
3DCoMPaT$^{++}$
Stats
"Our dataset comprises 10 million stylized 3D shapes rendered from 8 views, across 41 shape categories."
"We sample object-compatible combinations of part-material pairs to create 1000 styles per shape."
"For each image, camera parameters are provided."
Quotes
"We hope our work will help ease future research on compositional 3D Vision."
"Our dataset fills the gap in existing datasets by providing detailed part-level material information."