notesum.ai
Published at December 5Cubify Anything: Scaling Indoor 3D Object Detection
cs.CV
Released Date: December 5, 2024
Authors: Justin Lazarow1, David Griffiths1, Gefen Kohavi1, Francisco Crespo1, Afshin Dehghan1
Aff.: 1Apple
![[Uncaptioned image]](https://arxiv.org/html/2412.04458v1/x1.png)
| Traditional SUN RGB-D | Omni3D SUN RGB-D | CA-1M | ||||||||||
| Method | AP25 | AR25 | AP50 | AR50 | AP25 | AR25 | AP50 | AR50 | AP25 | AR25 | AP50 | AR50 |
| 3D point-based methods | ||||||||||||
| ImVoxelNet [15] (RGB only) | 41.0 | 74.9 | 13.5 | 29.0 | 14.4 | 39.0 | 2.5 | 8.8 | 10.1 | 22.8 | 2.3 | 6.3 |
| FCAF [14] | 63.5 | 94.2 | 47.0 | 72.5 | 27.1 | 56.5 | 15.6 | 30.4 | 29.3 | 49.5 | 11.2 | 22.6 |
| TR3D [16] | 66.2 | 93.6 | 49.7 | 72.6 | 27.1 | 64.2 | 15.2 | 30.9 | 22.0 | 51.9 | 4.4 | 20.0 |
| TR3D + FF [16] | 68.8 | 94.1 | 51.7 | 73.7 | 29.1 | 63.3 | 15.5 | 31.5 | 24.8 | 52.9 | 4.7 | 21.0 |
| 2D image-based methods | ||||||||||||
| Cube R-CNN [3] (RGB only) | - | - | - | - | 18.9 | 30.0 | 5.3 | 11.0 | 4.6 | 20.1 | 1.0 | 4.7 |
| CuTR (RGB only) | 45.9 | 75.3 | 17.0 | 40.2 | 21.5 | 40.4 | 6.9 | 16.9 | 13.5 | 35.4 | 2.4 | 12.9 |
| CuTR (RGB-D) | 59.4 | 87.2 | 34.0 | 56.4 | 30.3 | 60.2 | 13.6 | 29.0 | 40.9 | 62.3 | 12.7 | 29.1 |