3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics

Abstract

3D-FRONT is a new, large-scale, and comprehensive repository of synthetic indoor scenes with professionally and distinctively designed layouts, a large number (18,797) of rooms populated with 3D furniture objects that are stylistically compatible and endowed with high-quality textures. All freely available to the academic community and beyond.

What's Behind the Couch? Directed Ray Distance Functions (DRDF) for 3D Scene Reconstruction

“…Our primary dataset is Matterport3D [5], which satisfies all properties. We also evaluate on 3DFront [17] and ScanNet [10]. While 3DFront does not have capture holes, cutting it with a view frustum creates holes.…”

“…We then sample 13K/1K/1K images for train/val/test set. 3DFront [17]. This is a synthetic dataset of houses created by artists with underling hole-free 3D geometry.…”

“…This is a synthetic dataset of houses created by artists with underling hole-free 3D geometry. We collect 4K scenes from 3DFront [17] after removing scenes with missing annotations. We select 20 camera poses and filter for bad camera poses similar in Matterport3D [5].…”

“…To give a sense of scene statistics, we plot the frequency of the locations of the first 5 hits for each dataset (computed on 1M rays each) in Figure 5. 99% of ScanNet rays have 1 or 2 hits, while ≥24% of Matterport3D [5] and 3DFront [17] rays have more than 2 hits.…”

“…We compare our proposed approach, DRDF, with other distance functions and ways of obtaining a reconstruction of the scene, such as Layered Depth Images (LDI) [44]. Our experiments ( §4) compare the approaches on Matterport3D [5], 3DFront [17], and ScanNet [10]. Our evaluations show that the DRDF is substantially better at recovering the full 3D scene across a variety of metrics, including both visible and invisible regions.…”