Closed yxchng closed 3 weeks ago
I am closing this for now. Mask proposal is always a bottleneck for 3D open-world instance segmentation. Recent work, such as Segment3D, may provide a better solution for mask proposal, which we will also look into in the future.
Best, Zhening
Hi, thank you for your interest in our work. (Apologies for the super delay in response; I've just returned from a break.
The class-agnostic Mask3D is trained by modifying the following:
1) Removing the classification loss and adding IOU loss. 2) Removing the class label in Hungarian matching.
For training, please navigate to this folder and use this config file:
We suggest also referring to Mask3D for more information on data processing as well as environment setup. They have more comprehensive instructions there on how to train Mask3D. We adopted the easy setup for quick inference.