/INS-Conv

INS-Conv: Incremental Sparse Convolution for Online 3D Segmentation (CVPR 2022)

Primary LanguageCudaOtherNOASSERTION

INS-Conv: Incremental Sparse Convolution for Online 3D segmentation

This is the incremental sparse convolution library implemented based on SparseConvNet and Live Semantic 3D Perception for Immersive Augmented Reality. The later describes a more efficient GPU implementation of the original submanifold sparse convolution. Our method supports incremental computing of sparse convolution, including SSC, convolution/deconvolution, BN, IO, and residual structure, etc.

Environment setup

Preliminary Requirements:

  • Ubuntu 16.04
  • CUDA 9.0

Install

conda env create -f p1.yml

sh all_build.sh

Demo

For training, you could train an arbitary model using the original sparseconvnet.

For incremental inference, demo.py gives an example of the INS-Conv library.

We also provide the code for the online 3D semantic instance segmentation demo as in our video, you can download by the following link: https://drive.google.com/file/d/1sYpMFc1dVXZSZEDhfqQZbMoabiZZikuI/view?usp=sharing