Skip to content

doobiusP/TRINIT_firstTry_ML

Repository files navigation

Faster R-CNN Implementation by firstTry

Repository for ML Track Problem Statement 1 of Tri-NIT Hackathon 2024 by the team 'firstTry'

Link to Video

Members:

  • Aaryan Patil
  • Vedanth Nanesha
  • Alen Basil
  • Tanay Shekokar

Structure of the Repository:

The repository contains an XML parser and 2 python files that have been used to data-augment the existing dataset to ensure all images are 600x600 in size whilst also adjusting their corresponding bounding box coordinates. Augmented files are already provided for in the following kaggle link to the merged dataset.

The model is presented as 2 .ipynb files, 1 being the file used for training and visualising progress of the Faster R-CNN during training and the other being a file containing the minimal dependencies required to use the model for testing. Both .ipynb files contain markdown for easier navigation.

Pretrained parameters are uploaded as 'FasterRCNN_Params.pth' which can be loaded into the model in the testing .ipynb file.

Using the Repository:

To test the model, for which we have utilised transfer learning principles with RESNET50 base architecture, be it locally or on a site such as kaggle, provide the relative path inside your dataset to a folder that contains all testing images that should be classified. The testing code will display the first 20 images alongside their generated bounding boxes and associated labels.

To train the model using kaggle, connect to our publicly available dataset (link above) to any notebook and press 'Run All' to begin training. Statistics for every set of iterations will be printed in the log for tracking.

Alternatively, the original rough code for training Faster R-CNN exists in this public kaggle notebook . Accordingly, the original rough code for testing the Faster R-CNN exists in this public kaggle notebook .

Sample test images:

Sample Output Image Sample Output Image Sample Output Image

Key/Legend:

D00 - Longitudinal Crack

D10 - Transverse Crack

D20 - Alligator Crack

D40 - Pothole

D44 - Certain Road Markings like Zebra Crossings

D50 - Objects like Manholes

BG - Background for Images without Bounding Boxes

About

Repository of code for ML Track Problem Statement 1 of Tri-NIT Hackathon 2024 for the team 'firstTry'

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •