pranay731 / ros-semantic-segmentation-pytorch

Pytorch implementation of Semantic Segmentation in ROS on MIT ADE20K dataset based on semantic-segmentation-pytorch by CSAIL

Semantic Segmentation for ROS in PyTorch

Based on Sematic Segmentation on MIT ADE20K dataset in PyTorch by CSAIL-Vision

This is a ROS implementation for Semantic Segmentation, build upon semantic-segmentation-pytorch by MIT CSAIL Computer Vision Lab.

The implementation is in Python3 over PyTorch GPU, and holds training and evaluation parts of the parent project. "README_CSAIL.md" should be refered for more info over training.

Environment

The code is tested under the following configuration.

• Ubuntu 18.04 LTS
• CUDA=10.2
• PyTorch>=1.5.1
• ROS Melodic

1. Installation

A. Prerequisities

• Install ROS by following the official ROS website.

• Install dependencies:

sudo apt-get install python-catkin-tools git
# install Python3 dependencies
pip3 install rospkg
pip3 install numpy scipy
pip3 install pytorch>=0.4.1 torchvision
pip3 install yacs tqdm

B. Setup

Using catkin:

# Setup catkin workspace
mkdir -p ~/catkin_ws/src
cd ~/catkin_ws/
catkin init

# Add workspace to bashrc.
echo 'source ~/catkin_ws/devel/setup.bash' >> ~/.bashrc

# Clone repo
cd ~/catkin_ws/src
git clone https://github.com/pranay731/ros-semantic-segmentation-pytorch.git

sudo chmod +x ros-semantic-segmentation-pytorch/scripts/run_semantic_segmentation

# Build 
cd ..
catkin build

# Refresh workspace
source ~/catkin_ws/devel/setup.bash

A number of models are available. Default is resnet50dilated for encoder and ppm_deepsup for decoder. Download pretrained checkpoints for models from CSAIL Website and copy them to src/ros-semantic-segmentation-pytorch/ckpt/{modelname}/

2. Usage

A. Inference

Launch file :

• Input topic -> raw image rgb topic.
• gpu_id -> set gpu id for multiple gpu system else 0.
• cfg_filepath -> path to config file. Should be according to model.
• model_ckpt_dir -> path to directory containing downloaded checkpoints.

Configuration file contains option "imgSizes" which takes a tuple of heights, over which input image is resized for inference. This can be tweaked according to GPU capablities.

roslaunch semantic_segmentation_ros semantic_segmentation.launch

B. Train and Eval

Follow "README_CSAIL.md" for Training and Evaluation. Train and Eval scripts are present in src/ros-semantic-segmentation-pytorch/scripts directory.