Oxford-IIIT Pet Dataset - Instance Segmentation

• Notebooks
• Quick start
• Useful commands
• General info
• Folder structure
• Main Components
• Status
• Credits

1. Data analysis here.
2. Convert a custom dataset to COCO format here.
3. When making a custom dataset check if your data is compatible with the COCO format here.
4. Create a config to train the model here.
5. Demo training here.

Nice virtualenv tutorial here

pip3 install --upgrade pip

which python3.8
mkvirtualenv -p <path to python3> <name>
workon <name>

For example:

mkvirtualenv -p /usr/bin/python3.8 Febrin
workon Febrin

Ensure that you have properly installed all the required packages:

• CUDA 11.0
• GCC 7.5
• torch 1.7.1+cu110
• torchvision 0.8.2+cu110
• Comet ML 3.9.0
• OpenCV: 4.1.2
• MMCV: 1.3.1
• MMDetection: 2.11.0+187774b

In our case we use torch 1.7.1 and Cuda 11.0

$ nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.102.04   Driver Version: 450.102.04   CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+

Install correct pytorch

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html

Install mmdetection with mmcv

rm -rf mmdetection
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -e .

So this is matching mmcv

pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu110/torch1.7.0/index.html

Download dataset and models

python download.py

Test Your own images

python inference.py image.jpg

Our goal is to create an instance segmentation model based on the Oxford-IIIT Pet Dataset.

Instance segmentation is a task in Computer Vision that aims to identify each instance of each object within the image at the pixel level. In our case, each image contains a single object, which means our goal is to predict which pixels belong to the object and which to the background as well as predict the object's class.

Our dataset contains ~200 images for each from 37 classes.

Near 2/3 of those classes are dog breeds, the rest of them are cat breeds.

The images have different sizes and rations, therefore they need to be resized. There's a high variance in background colors and the amount of light on the pictures.

We will build our model using MMDetection - an open-source object detection toolbox based on PyTorch. We will experiment with different types of backbones, for example, ResNet, ResNext, or VGG.

├── /model                 - folder for keeping models
│   └── download_model.py  - script for downloading dataset
│
├── /trainer               - training scripts
│   └── train.py
|   └── test.py
│   
├──  /mains                - main files responsible for the whole pipeline
│    └── main.py 
|
├──  /figures              - figures generated during analysis and used in README
│    └── ...
│ 
├──  /notebooks            - notebook files created for tests on Colaboratory
│    ├── /Convert_to_COCO_format.ipynb
│    ├── /Create_config.ipynb
│    ├── /Demo training.ipynb
│    ├── /Demo.ipynb
│    ├── /Pycoco-test.ipynb
│    └── /Data_analysis.ipynb
| 
├──  /dataset              - things related to the dataset
│    ├── /train            - train datapoints and labels
│    ├── /test             - test datapoints and labels
│    ├── /valid            - valid datapoints and labels
│    └── divide_dataset.py - script that divides the dataset into /train/test/valid
│
└── /utils 
     ├── logger.py
     └── ...

This template also supports reporting to Comet.ml which allows you to see all your hyper-params, metrics, graphs, dependencies and more including real-time metric.

Here's how it looks after you start training:

You can also link your Github repository to your comet.ml project for full version control. Here's a live page showing the example from this repo

Project started: 09.03.2021

1. Week 1:

   • Created the repository
   • Analyzed the dataset
   • Investigated data storing options (GCS) and model training (GCP machine)
   • Created a Kanban board where we track the tasks and progress.

2. Week 2:

   • Downloaded the data and created a dataloader
   • Started to familiarize ourselves with mmdet environment
   • Tested CometML and played with it
   • Created requirements.txt

3. Week 3:

   • Created our own CometML logger hook and integrated it with their API
   • Created a proper DataLoader
   • Started writing a mmdet notebook that will transition to train.py soon
   • Read some papers i.e. Feature Pyramid Network

4. Week 4:

   • Validated our coco using: https://github.com/cocodataset/cocoapi/tree/master/PythonAPI/pycocotools
   • Cleaned the repo and README
   • Train on the second model to show that COMET ML

5. Week 5:

   • Moved our code from Colab files to Python scripts
   • Created Docerfile to be used as an environment created inside Colab

Project ended: 27.04.2021

• @mdragula
• @MatMarkiewicz
• @TheFebrin