2021-04-21 The augmented 3D-FUTURE dataset, including 20,240 indoor images and 16,563 (+6571) 3D furniture models, has been released.
2020-04-03 Release of training and validation data for the Alibaba 3D Artificial Challenge 2020.
High-quality Shapes, Informative Textures, Rich Attributes
The 3D shapes offered by public 3D benchmarks may show two imperfections. Firstly, most of these 3D CAD models (for furniture) are both with fewer details and low informative textures since they are collected online. Secondly, there are no diverse professional attributes for their furniture shapes. In contrast, 3D-FUTURE provides high-quality 3D furniture with rich details in various styles, including European furniture that often contains intricate carvings. Furthermore, each 3D shape in 3D-FUTURE is assigned to an informative texture and different attribute labels. We believe these features can potentially facilitate innovative research on high-quality 3D shape understanding and generation.
Realistic Renderings, Real 2D-3D Alignment
There are no well-organized benchmarks that provide realistic synthetic indoor images. 3D-FUTURE fill the blank by rendering 20,240 photo-realistic synthetic images across 5,000 scenes via one of the most advanced industrial 3D renders (V-Ray). These indoor scenes are reviewed by professional designers. Besides, existing benchmarks only provide pseudo 2D-3D alignment annotations. Namely, they manually choose a roughly matched 3D CAD model from public 3D shape benchmarks according to the object contained in the image. Annotators thus may largely ignore some local shape details. As a result, these benchmarks offer less matched 3D shape and 2D image pairs. This is not sufficient to support data-driven studies such as high-quality 3D reconstruction and high-accuracy 3D shape retrieval. Luckily, the 9,992 3D shapes in 3D-FUTURE exactly match objects contained in the rendered images.