Skip to content

Latest commit

 

History

History
53 lines (36 loc) · 2.68 KB

README.md

File metadata and controls

53 lines (36 loc) · 2.68 KB
Raw

Point-GCC: Universal Self-supervised 3D Scene Pre-training via Geometry-Color Contrast

Installation

This implementation is based on mmdetection3d v1.0 framework. Please refer to the original installation guide getting_started.md, including MinkowskiEngine installation.

Getting Started

Please refer to the original guide getting_started.md for basic usage examples and data preparation for scannet, sunrgbd, and s3dis.

Usage

Pre-train

3DCNN

To start pre-training, run train with 3DCNN backbone such as for TD3D:

./tools/dist_train.sh configs/point_gcc_3dcnn/td3d_scannet.py 4 --no-validate

PointNet

To start pre-training, run train with PointNet backbone such as for VoteNet:

./tools/dist_train.sh configs/point_gcc_pointnet/votenet-scannet.py 4 --no-validate

Fine-tuning

VoteNet

To start fine-tuning, modify the load_from field to your pretrain model path, and run train with VoteNet for object detection:

./tools/dist_train.sh configs/votenet/votenet_8x8_scannet-3d-18class-fine.py 8

GroupFree3D

To start fine-tuning, modify the load_from field to your pretrain model path, and run train with GroupFree3D for object detection:

./tools/dist_train.sh configs/groupfree3d/groupfree3d_8x4_scannet-3d-18class-L6-O256-fine.py 4

PointNet++(SSG)

To start fine-tuning, modify the load_from field to your pretrain model path, and run train with PointNet++(SSG) for semantic segmentation:

./tools/dist_train.sh configs/pointnet2/pointnet2_ssg_16x2_cosine_200e_scannet_seg-3d-20class.py 2

TR3D and TD3D

To start fine-tuning with TR3D and TD3D, please follow the official repo and modify the load_from field to your pretrain model path in corresponding config file.