PyTorch for Semantic Segmentation

4 min readOct 15, 2020

This repository contains some models for semantic segmentation and the pipeline of training and testing models, implemented in PyTorch

Models

Vanilla FCN: FCN32, FCN16, FCN8, in the versions of VGG, ResNet and DenseNet respectively (Fully convolutional networks for semantic segmentation)
U-Net (U-net: Convolutional networks for biomedical image segmentation)
SegNet (Segnet: A deep convolutional encoder-decoder architecture for image segmentation)
PSPNet (Pyramid scene parsing network)
GCN (Large Kernel Matters)
DUC, HDC (understanding convolution for semantic segmentation)
Mask-RCNN (paper, code from FAIR, code PyTorch)

Requirement

PyTorch 0.2.0
TensorBoard for PyTorch. Here to install
Some other libraries (find what you miss when running the code :-P)

Preparation

Go to models directory and set the path of pretrained models in config.py
Go to datasets directory and do following the README

TODO

I’m going to implement The Image Segmentation Paper Top10 Net in PyTorch firstly.

DeepLab v3
RefineNet
More dataset (e.g. ADE)

A CSAE: pytorch-mask-rcnn

This is a Pytorch implementation of Mask R-CNN that is in large parts based on Matterport’s Mask_RCNN. Matterport’s repository is an implementation on Keras and TensorFlow. The following parts of the README are excerpts from the Matterport README. Details on the requirements, training on MS COCO and detection results for this repository can be found at the end of the document.

The Mask R-CNN model generates bounding boxes and segmentation masks for each instance of an object in the image. It’s based on Feature Pyramid Network (FPN) and a ResNet101 backbone.

The next four images visualize different stages in the detection pipeline:

1. Anchor sorting and filtering

The Region Proposal Network proposes bounding boxes that are likely to belong to an object. Positive and negative anchors along with anchor box refinement are visualized.

2. Bounding Box Refinement

This is an example of final detection boxes (dotted lines) and the refinement applied to them (solid lines) in the second stage.

3. Mask Generation

Examples of generated masks. These then get scaled and placed on the image in the right location.

4. Composing the different pieces into a final result

Requirements

Python 3
Pytorch 0.3
matplotlib, scipy, skimage, h5py

Installation

Clone this repository.

git clone https://github.com/multimodallearning/pytorch-mask-rcnn.git

2. We use functions from two more repositories that need to be build with the right --arch option for cuda support. The two functions are Non-Maximum Suppression from ruotianluo's pytorch-faster-rcnn repository and longcw's RoiAlign.

GPU arch

TitanX sm_52

GTX 960M sm_50

GTX 1070 sm_61

GTX 1080 (Ti) sm_61

[1] cd nms/src/cuda/  
[2]nvcc -c -o nms_kernel.cu.o nms_kernel.cu -x cu -Xcompiler -fPIC -[3]arch=[arch]  
[4]cd ../../  
[5]python build.py  
[6]cd ../   
[7]cd roialign/roi_align/src/cuda/  
[8]nvcc -c -o crop_and_resize_kernel.cu.o crop_and_resize_kernel.cu -x cu -Xcompiler -fPIC -arch=[arch]  
[9]cd ../../  
[10]python build.py  
[11]cd ../../

3. As we use the COCO dataset install the Python COCO API and create a symlink.

ln -s /path/to/coco/cocoapi/PythonAPI/pycocotools/ pycocotools

4. Download the pretrained models on COCO and ImageNet from Google Drive.

Demo

To test your installation simply run the demo with

python demo.py

It works on CPU or GPU and the result should look like this:

Training on COCO

Training and evaluation code is in coco.py. You can run it from the command line as such:

# Train a new model starting from pre-trained COCO weights
python coco.py train --dataset=/path/to/coco/ --model=coco# Train a new model starting from ImageNet weights
python coco.py train --dataset=/path/to/coco/ --model=imagenet# Continue training a model that you had trained earlier
python coco.py train --dataset=/path/to/coco/ --model=/path/to/weights.h5# Continue training the last model you trained. This will find
# the last trained weights in the model directory.
python coco.py train --dataset=/path/to/coco/ --model=last

If you have not yet downloaded the COCO dataset you should run the command with the download option set, e.g.:

# Train a new model starting from pre-trained COCO weights
python coco.py train --dataset=/path/to/coco/ --model=coco --download=true

You can also run the COCO evaluation code with:

# Run COCO evaluation on the last trained model
python coco.py evaluate --dataset=/path/to/coco/ --model=last

The training schedule, learning rate, and other parameters can be set in coco.py.

Results

COCO results for bounding box and segmentation are reported based on training with the default configuration and backbone initialized with pretrained ImageNet weights. Used metric is AP on IoU=0.50:0.95.

from scratchconverted from kerasMatterport’s Mask_RCNNMask R-CNN paperbboxt.b.a.0.3470.3470.382segmt.b.a.0.2960.2960.354

About

No description, website, or topics provided.

Resources

Readme

License

View license