Skip to content

Latest commit

 

History

History
210 lines (141 loc) · 6.52 KB

README.md

File metadata and controls

210 lines (141 loc) · 6.52 KB
Raw

RGBD_Semantic_Segmentation_PyTorch

license PyTorch-1.0.0

Implement some state-of-the-art methods of RGBD Semantic Segmentation task in PyTorch.

Currently, we provide code of:

  • SA-Gate, ECCV 2020 [arXiv]
  • Malleable 2.5D Convolution, ECCV 2020 [arXiv]

News

  • 2020/08/16

Official code release for the paper Malleable 2.5D Convolution: Learning Receptive Fields along the Depth-axis for RGB-D Scene Parsing, ECCV 2020. [arXiv], [code]

Thanks aurora95 for his open source code!

  • 2020/07/20

Official code release for the paper Bi-directional Cross-Modality Feature Propagation with Separation-and-Aggregation Gate for RGB-D Semantic Segmentation, ECCV 2020. [arXiv], [code]

​

Main Results

Results on NYU Depth V2 Test Set with Multi-scale Inference

Method mIoU (%)
3DGNN 43.1
ACNet 48.3
RDFNet-101 49.1
PADNet 50.2
PAP 50.4
Malleable 2.5D 50.9
SA-Gate 52.4

Results on CityScapes Test Set with Multi-scale Inference (out method uses output stride=16 and does not use coarse-labeled data)

Method mIoU (%)
PADNet 80.3
DANet 81.5
GALD 81.8
ACFNet 81.8
SA-Gate 82.8

For more details, please refer to our paper.

​

Directory Tree

Your directory tree should look like this:

./
|-- furnace
|-- model
|-- DATA
-- |-- pytorch-weight
-- |-- NYUDepthv2
   |   |-- ColoredLabel
   |   |-- Depth
   |   |-- HHA
   |   |-- Label
   |   |-- RGB
   |   |-- test.txt
   |   |-- train.txt

Installation

The code is developed using Python 3.6 with PyTorch 1.0.0. The code is developed and tested using 4 or 8 NVIDIA TITAN V GPU cards. You can change the input size (image_height and image_width) or batch_size in the config.py according to your available resources.

  1. Clone this repo.

    $ git clone https://github.com/charlesCXK/RGBD_Semantic_Segmentation_PyTorch.git
$ cd RGBD_Semantic_Segmentation_PyTorch

  2. Install dependencies.

    (1) Create a conda environment:

    $ conda env create -f rgbd.yaml
$ conda activate rgbd

    (2) Install apex 0.1(needs CUDA)

    $ cd ./furnace/apex
$ python setup.py install --cpp_ext --cuda_ext

​

Data preparation

Pretrained ResNet-101

Please download the pretrained ResNet-101 and then put it into ./DATA/pytorch-weight.

Source Link
BaiDu Cloud Link: https://pan.baidu.com/s/1Zc_ed9zdgzHiIkARp2tCcw Password: f3ew
Google Drive https://drive.google.com/drive/folders/1_1HpmoCsshNCMQdXhSNOq8Y-deIDcbKS?usp=sharing

NYU Depth V2

You could download the official NYU Depth V2 data here. After downloading the official data, you should modify them according to the structure of directories we provide. We also provide the processed data. We will delete the link at any time if the owner of NYU Depth V2 requests.

Source Link
BaiDu Cloud Link: https://pan.baidu.com/s/1iU8m20Jv9shG_wEvwpwSOQ Password: 27uj
Google Drive https://drive.google.com/drive/folders/1_1HpmoCsshNCMQdXhSNOq8Y-deIDcbKS?usp=sharing

How to generate HHA maps?

If you want to generate HHA maps from Depth maps, please refer to https://github.com/charlesCXK/Depth2HHA-python.

​

Training and Inference

We just take SA-Gate as an example. You could run other models in a similar way.

Training

Training on NYU Depth V2:

$ cd ./model/SA-Gate.nyu
$ export NGPUS=8
$ python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py

If you only have 4 GPU cards, you could:

$ cd ./model/SA-Gate.nyu.432
$ export NGPUS=4
$ python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py
  • Note that the only difference between SA-Gate.nyu/ and SA-Gate.nyu.432/ is the training/inference image crop size.
  • The tensorboard file is saved in log/tb/ directory.

Inference

Inference on NYU Depth V2:

$ cd ./model/SA-Gate.nyu
$ python eval.py -e 300-400 -d 0-7 --save_path results
  • Here, 300-400 means we evaluate on checkpoints whose ID is in [300, 400], such as epoch-300.pth, epoch-310.pth, etc.
  • The segmentation predictions will be saved in results/ and results_color/, the former stores the original predictions and the latter stores colored version. Performance in mIoU will be written to log/*.log. You will expect ~51.4% mIoU in SA-Gate.nyu and ~51.5% mIoU in SA-Gate.nyu.432. (single scale inference with no flip)
  • For multi-scale and flip inference, please set C.eval_flip = True and C.eval_scale_array = [1, 0.75, 1.25] in the config.py. Different eval_scale_array may have different performances.

​

Citation

Please consider citing this project in your publications if it helps your research.

@inproceedings{chen2020-SAGate,
  title={Bi-directional Cross-Modality Feature Propagation with Separation-and-Aggregation Gate for RGB-D Semantic Segmentation},
  author={Chen, Xiaokang and Lin, Kwan-Yee and Wang, Jingbo and Wu, Wayne and Qian, Chen and Li, Hongsheng and Zeng, Gang},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2020}
}
@inproceedings{xing2020-melleable,
  title={Malleable 2.5D Convolution: Learning Receptive Fields along the Depth-axis for RGB-D Scene Parsing
},
  author={Xing, Yajie and Wang, Jingbo and Zeng, Gang},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2020}
}

Acknowledgement

Thanks TorchSeg for their excellent project!

​

TODO

  • More encoders such as HRNet.
  • Code and data for Cityscapes.
  • More RGBD Semantic Segmentation models