The 3D semantic instance prediction task involves detecting and segmenting the object in an 3D scan mesh.

Evaluation and metrics

Our evaluation ranks all methods according to the average precision for each class. We report the mean average precision AP at overlap 0.25 (AP 25%), overlap 0.5 (AP 50%), and over overlaps in the range [0.5:0.95:0.05] (AP). Note that multiple predictions of the same ground truth instance are penalized as false positives.



This table lists the benchmark results for the 3D object detection with limited reconstructions scenario.




Method Infoavg ap 25%bathtubbedbookshelfcabinetchaircountercurtaindeskdoorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwindow
sort bysorted bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort by
WS3D_LR_ODpermissive0.550 10.867 10.892 10.728 10.346 10.837 10.384 10.305 10.539 10.444 10.425 10.155 10.464 10.294 10.507 10.865 10.486 10.897 10.461 1
Kangcheng Liu: WS3D: Weakly Supervised 3D Scene Segmentation with Region-Level Boundary Awareness and Instance Discrimination. European Conference on Computer Vision (ECCV), 2022
CSC_LR_DET0.365 20.722 20.717 20.379 20.238 20.747 30.089 20.131 30.250 40.179 30.259 20.006 40.391 20.090 40.300 20.837 20.197 30.877 20.153 3
Scratch_LR_DET0.215 40.667 30.238 40.178 40.079 40.577 40.012 40.022 40.251 30.104 40.027 40.007 30.013 40.229 30.218 30.498 40.156 40.564 40.035 4
PointContrast_LR_DET0.361 30.667 30.695 30.358 30.156 30.757 20.070 30.186 20.432 20.253 20.225 30.016 20.322 30.269 20.156 40.730 30.199 20.793 30.210 2