Faster-RCNN-based-Object-Detection

Objective This project was created to find the spinal disorder Stenosis.

This repository walks through the Object detection process including the custom dataset preparation, training with model congifuration setup and inference of the trained model.

Requirements

1. Python 3.6
2. Numpy
3. Opencv
4. Pandas
5. Matplotlib
6. Tensorflow_gpu==1.15
7. Lvis

Note

Download the pretrained model(frozen graph) and the image dataset from the following link : https://drive.google.com/drive/folders/1ciCiNYvEpeCRIeP4QDzF5qIvnUCUj7UD?usp=sharing

Dataset Preparation

With few data augmentation techniques such as rotate, flip, blur and translate, incereased the size of the dataset. Further details are in data_prep_img_process.py

Annotation Tool : MakeSense http://makesense.ai/ Annotated data is stored as a csv file and from that source csv file, the train.csv and test.csv are genearated. With the data available on the test.csv and train.csv, the image dataset is split into test_set and train_set.

Training

Initially Tensorflow Object Detection API module is installed with all necessary packages.(tensorflow_gpu==1.15)

ref : https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/install.html#tensorflow-object-detection-api-installation

Generate the tfrecords for the test and training set using generate_tf_records.py

Model Architecture

For quick training and better results, this object detection module was trained on Tensorflow's faster_rcnn_inception_v2 model.

Hyper Parameters

1.num-steps = 5000 2.eval-steps=50 3.batch-size = 12 4.initial_learning_rate = 0.0002 5.num-classes = 1

Model Performance Evaluation

Displayed using Tensorboard on the python notebook.