search

Jun-CEN / Open-world-3D-semantic-segmentation

[ECCV 2022] Open-world Semantic Segmentation for LIDAR Point Clouds
67 stars 10 forks source link

The category selection of the sythesis from old classes #7

Closed yifliu3 closed 1 year ago

yifliu3 commented 1 year ago

Thanks for sharing your great work!

I'm a little confused about the category selection of the synthesis procedure. As described in your paper, the old classes used to synthesize novel objects are car for SemanticKITTI and car, bus, and truck for nuScenes, so what is the basis of this selection strategy?

Jun-CEN commented 1 year ago

Hi,

The only requirement of old classes used for synthesize novel objects is that the chosen class should belong to 'thing' class not 'stuff', so that we have the instance IDs of objects and use them for synthesize. For different datasets, we do not have additional reasons or bias, we just pick them.