YOLOV1 Training Pipeline on the PASCAL VOC Data using the PyTorch Framework

This repository contains YOLOV1 training pipeline on the PASCAL VOC 2007 and 2012 data using the PyTorch framework. A few points:

Note: Use this repository/code for any project/learning. A simple reference/attribution to the repository is enough. Completely OPEN for any use.

Steps to Train

First of All, Download and Extract the Data (Download to any directory of your choice)

• Download VOC 2007 trainval from here.

  • Or wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar in terminal.

• Download VOC 2007 test from here.

  • Or wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar in terminal.

• Download the VOC 2012 trainval from here.

  • Or wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar

Execute the following commands in the directory where the downloaded .tar files are present.

tar xf VOCtrainval_06-Nov-2007.tar

tar xf VOCtest_06-Nov-2007.tar 

tar xf VOCtrainval_11-May-2012.tar

Prepare the Text Files and Labels

• Then execute prepare_data.py with the correct path argument to the VOCdevkit parent directory. Simply, the next folder after the path in the arugment should be VOCdevkit The following is an example:

  python prepare_data.py --path my_pacal_voc_data
  

  In the above command my_pacal_voc_data should contain the VOCdevkit directory for the dataset.

• Then execute prepare_text_labels.py with correct command line argument paths to the VOC 2007 and 2012 Annotations directory. See text_labels.py to know what the exact path should be. Example command:

  python prepare_text_labels.py --annotations-2007 VOCdevkit/VOC2007/Annotations/ --annotations-2012 VOCdevkit/VOC2012/Annotations/
  

Train

• Execute python train.py.

References

• YOLOv1 paper: https://arxiv.org/pdf/1506.02640v5.pdf
• Some dataset preparation ideas borrowed from: https://github.com/aladdinpersson/Machine-Learning-Collection/blob/master/ML/Pytorch/object_detection/YOLO/dataset.py