pytorch-deeplab-xception

Differences to parent repository

• Load RGB-D datasets Cityscapes, COCO, SUNRGBD, and SceneNetRGBD
• Optional RGB-D network input using 4th channel in first convolutional layer
• YACS configuration files

Introduction

This is a PyTorch(1.4) implementation of DeepLab-V3-Plus. It can use Modified Aligned Xception and ResNet as backbone.

Installation

The code was tested with Anaconda and Python 3.8. After installing the Anaconda environment:

0. Clone the repo:

   git clone https://github.com/jfzhang95/pytorch-deeplab-xception.git
   cd pytorch-deeplab-xception

1. Install dependencies:

   For PyTorch dependency, see pytorch.org for more details.

   Other dependencies:

   conda install matplotlib pillow tqdm protobuf tabulate scipy numpy pandas
   pip install tensorboardX yacs

   Coco tools

   conda install -c conda-forge pycocotools scikit-image

2. Install dataloader repo

   This is a seperate repo for use in different projects. For more details, see the repo README

   git clone https://github.com/crmauceri/dataset_loaders.git
   cd dataset_loaders/datasets
   make
   cd ../..
   pip install -e .

3. Compile CUDA code

   The depth-aware convolution and depth-aware average pooling operations are under folder modeling/backbone/ops/, to build them, simply use python setup.py install to compile.

   cd modeling/backbone/ops/depthconv/
   python setup.py install
   
   cd ../depthavgpooling/
   python setup.py install
   
   cd ../../../..

4. Compile SceneNetRGBD protobuf files

   cd dataloaders/datasets
   make

5. Install as module:

   cd $root
   pip install -e .

Model Zoo

RGB Models

From parent repo,

Backbone	Dataset	mIoU in val	Pretrained Model
ResNet (pretrained)	Pascal VOC 2012	78.43%	google drive

New models,

Backbone	Dataset	mIoU in val	Pretrained Model
ResNet (pretrained)	COCO		google drive
ResNet	COCO		google drive

RGBD Models

Backbone	Dataset	mIoU in val	Pretrained Model
ResNet	COCO + Synthetic Depth		google drive
ResNet	Cityscapes		google drive
ResNet	SceneNetRGBD		google drive

Training

Follow steps below to train your model:

0. Configuration files use YACS. Full list of defaults is in deeplab3/config/defaults.py. Configurations used to train model zoo are in configs/. You must set DATASET.ROOT to match your dataset location, otherwise, you can change as many or as few parameters as you desire.

1. Run training script as To train deeplabv3+ using COCO dataset and ResNet as backbone:

   python deeplab3/train.py <config_file> <optional_parameter_overrides>

   For example, to train deeplabv3+ using SUNRGBD dataset and ResNet as backbone:

   python deeplab3/train.py configs/sunrgbd.yaml

Visualization

Jupyter notebook COCO_Data_Browser.ipynb is provided for visualizing data and trained network results.

Tensorboard can also be used to visualize loss and accuracy metrics during training.

tensorboard serve --log=run/<output_path>

Acknowledgement

PyTorch-Encoding

Synchronized-BatchNorm-PyTorch

drn