3D Semantic Instance Results - ScanNet Benchmark

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

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.

Method	avg ap 50%	bathtub	bed	bookshelf	cabinet	chair	counter	curtain	desk	door	otherfurniture	picture	refrigerator	shower curtain	sink	sofa	table	toilet	window

R-PointNet	0.306 1	0.500 4	0.405 4	0.311 2	0.348 1	0.589 1	0.054 1	0.068 4	0.126 2	0.283 1	0.290 1	0.028 3	0.219 1	0.214 3	0.331 1	0.396 4	0.275 1	0.821 2	0.245 1

3D-SIS	0.271 2	0.528 3	0.558 3	0.241 3	0.272 2	0.488 2	0.000 5	0.093 2	0.167 1	0.179 2	0.231 2	0.040 2	0.182 2	0.286 2	0.168 2	0.546 2	0.155 4	0.555 4	0.185 2

3D-BEVIS	0.248 3	0.667 1	0.566 2	0.076 5	0.035 5	0.394 3	0.027 3	0.035 5	0.098 3	0.099 4	0.030 4	0.025 4	0.098 4	0.375 1	0.126 3	0.604 1	0.181 2	0.854 1	0.171 3

MTML	0.212 4	0.667 1	0.614 1	0.337 1	0.027 6	0.390 4	0.000 5	0.118 1	0.001 4	0.100 3	0.028 5	0.000 6	0.167 3	0.143 4	0.046 5	0.500 3	0.105 5	0.570 3	0.003 6

Sgpn_scannet	0.143 5	0.208 6	0.390 5	0.169 4	0.065 3	0.275 5	0.029 2	0.069 3	0.000 5	0.087 5	0.043 3	0.014 5	0.027 6	0.000 5	0.112 4	0.351 5	0.168 3	0.438 5	0.138 4

MaskRCNN 2d->3d Proj	0.058 6	0.333 5	0.002 6	0.000 6	0.053 4	0.002 6	0.002 4	0.021 6	0.000 5	0.045 6	0.024 6	0.238 1	0.065 5	0.000 5	0.014 6	0.107 6	0.020 6	0.110 6	0.006 5

3D Semantic instance benchmark