Object_Detection

Description

This is a custom pipeline dedicated to solving object detection problems and is derived from Quicksetup-ai which is based on lightning-hydra, a flexible template as a quick setup for deep learning projects in research. The pipeline comes with preloaded Faster RCNN architecture and can be easily extended by adding other SOTA architectures in Computer vision.

Quickstart

1. Install dependencies

1.1 Installing dependencies on your local machine (with CPU or GPU) and creating conda environment.

# clone project
git clone https://github.com/shvardhan1994/object_detection.git
cd object_detection

# [OPTIONAL] create conda environment
conda create -n venv python>=3.8
conda activate venv

# Ensure that the lines between 51 and 93 are not commented in the setup.cfg file.

# install requirements and "object_detection module"
pip install -e .

Run the organoid detection example

Note: This pipeline performs object detection on intestinal organoid images. The dataset is split into train, validation and test set. The dataset has to be saved in object_detection/data/orgaquant folder. The data folder should follow below structure.

This object detection template is customely designed to perform on the sample dataset acquired from the below research, Kassis, Timothy, et al. "OrgaQuant: human intestinal organoid localization and quantification using deep convolutional neural networks." Scientific reports 9.1 (2019): 12479. The opensource dataset can be download from here.

├── object_detection                             
│   ├── data                               
│       ├── orgaquant                             
│           ├── train
            ├── val
            ├── test
            ├── train_labels.csv
            ├── val_labels.csv
            ├── test_labels.csv
        ├── predictions
        ├── metrics

The .csv files contains the annotations for the bounding boxes of organoids present in the input images. The .csv file has below structure,

x1 | y1 | x2 | y2 | class_name | path

To run the training pipeline, simply run:

python scripts/train.py

By default, the training pipeline also runs the testing pipeline on the testset based on the best model checkpoint during training.

If you want to run just testing pipeline exclusively, make sure you give the correct checkpoint path in configs/test.yaml and then simply run:

python scripts/test.py

The predictions are saved by default in this path as a json file object_detection/scripts/data/orgaquant/predictions

To compute the evaluation metrics on the testset (this pipeline calculates Precisin-Recall values for different box score thresholds and saves them as an array in object_detection/data/orgaquant/metrics), specify the path to the predictions in .json format as mentioned previously in the inference script in object_detection/notebooks/inference.py, then simply run:

python notebooks/inference.py

• The experiments, evaluations, etc., are stored under the logs directory.
• The default experiments tracking system is mlflow. The mlruns directory is contained in logs. To view a user friendly view of the experiments, run:

# make sure you are inside logs (where mlruns is located)
mlflow ui --host 0000

• When evaluating (running test.py), make sure you give the correct checkpoint path in configs/test.yaml