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 semantic instance with limited annotations scenario.




Method Infoavg apbathtubbedbookshelfcabinetchaircountercurtaindeskdoorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwindow
sort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysort bysorted bysort bysort bysort bysort bysort bysort bysort by
Box2Mask_LA0.235 20.407 20.199 40.427 10.107 20.414 20.007 20.229 30.022 20.098 20.226 20.152 10.261 20.492 20.039 40.427 10.113 30.463 40.150 2
Julian Chibane, Francis Engelmann, Tuan Anh Tran, Gerard Pons-Moll: Box2Mask: Weakly Supervised 3D Semantic Instance Segmentation Using Bounding Boxes. ECCV 2022
WS3D_LA_Inspermissive0.336 10.630 10.386 10.242 20.173 10.575 10.007 10.550 10.050 10.164 10.232 10.127 20.424 10.614 10.198 10.420 20.314 10.771 10.164 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_LA_INS0.159 30.343 30.247 30.226 30.063 30.264 50.000 30.268 20.004 30.018 40.064 40.007 30.009 40.041 50.091 20.279 30.194 20.712 20.037 3
PointContrast_LA_INS0.124 40.037 50.109 50.085 40.052 40.287 30.000 30.219 40.001 40.040 30.104 30.002 40.074 30.211 30.055 30.239 40.036 50.661 30.020 4
Scratch_LA_INS0.089 50.306 40.319 20.025 50.035 50.274 40.000 30.157 50.001 50.006 50.031 50.000 50.000 50.095 40.017 50.172 50.077 40.094 50.002 5