This is the 《Multi-modal Arbitrary Rotation based Self-supervised Learning for RGB-D Semantic Segmentation》 official implementation code.

For the ArbRot operation, please refer to ArbRot\mmseg\datasets\pipelines\transforms_img_hha.py RotateCircleImgAndHHA function

For the Multi-modal Multi-task Self-supervised Learning, please refer to:
ArbRot\mmseg\models\segmentators\MoCoUnet0801.py for MoCo,
ArbRot\mmseg\models\segmentators\SimCLRUnet1026 for SimCLR,
ArbRot\mmseg\models\segmentators\BYOLUnet1026 for BYOL,
ArbRot\mmseg\models\segmentators\SimSiamUnet0623 for SimSiam

1. Requirements

• Linux
• Python 3.6+
• PyTorch 1.7.0 or higher
• CUDA 10.0 or higher

We have tested the following versions of OS and softwares:

• OS: Ubuntu 16.04.6 LTS
• CUDA: 10.0
• PyTorch 1.7.0
• Python 3.6.9

2. Install dependencies.

pip install -r requirements.txt

1. Download SUNRGBD v2 data HERE (SUNRGBD.zip, SUNRGBDMeta2DBB_v2.mat, SUNRGBDMeta3DBB_v2.mat) and the toolkits (SUNRGBDtoolbox.zip). Move all the downloaded files under OFFICIAL_SUNRGBD. Unzip the zip files.

2. Extract point clouds and annotations (class, v2 2D -- xmin,ymin,xmax,ymax, and 3D bounding boxes -- centroids, size, 2D heading) by running extract_split.m, extract_rgbd_data_v2.m and extract_rgbd_data_v1.m under the matlab folder.

We provide ready-made annotation files we generated offline for reference, which are provided by MMdetection3D. You can directly use these files for convenice.

Or download the converted dataset:

• NYU-V2

1. Config

   Edit config file in ./configs.

   1. Set data_root = path_to_dataset.

      E.g.,data_root = "data/sunrgbd/sunrgbd_trainval".

2. Pretrain Run

   1. Distributed training

   ./tools/dist_train.sh config_path gpu_num --no-validate

   E.g., self-supervised train the model on SUN RGB-D with 2 GPUs, please run:

   .tools/dist_train_moco1.sh configs/selfsup/moco/from_scratch/1027_marc_res50_8b8_lr_1e-3_300e_scratch_byol_loc_rot_rgd_dgr.py 2 --no-validate

   2. Non-distributed training

   python tools/train.py config_path  --no-validate

We transfer the obtained weights of our Multi-modal Arbitrary Rotation based Self-supervised Learning to DeepLabv3 and ShapeConv(https://github.com/hanchaoleng/ShapeConv) methods as initialization.

This repository is based on MMSegmentation,MMDetection3D, MMSelfsup, ShapeConv