This project aims to compare and evaluate the performance of transformer-based and traditional deep-learning object detection models on different image enhancement techniques.
The project is an extended and improved version of the final project presented for CS-577 Deep Learning course at Illinois Institute of Technology. The original project can be found here. This project was developed by me, Ignacio Gomez Valverde and Prashanth V.R. who made the a first version of the YOLO preprocessing and training algorithm.
The results can be found in the project's final report, included in the Anexes section under the LinkedIn publication link.
The Exclusively Dark (ExDark) dataset was used tu develop dis project. It contains the largest collection of natural low-light images taken in visible light to date, including object level annotation.
Preparation and preprocessing functions are provided in the utils folder.
- 3000 images for training - 250 per class
- 1800 images for validation - 150 per class
- 2563 images for testing - rest of the images per class
Some usefull custom made functions for the project that can be executed from the terminal (locating the user in this folder using the cd command) or from the data-preparation notebook.
For the transformer model, a cell is provided with the different enhancement options or "raw" if user just wants to train using default data.
Disclaimer: A resize and some othe preprocessing adjustments are made either if the user selects an option or leaves the "raw" version of the images.
- Ignacio Gomez Valverde - LinkedIn Profile
- End-to-End Detection with Transformers
- Detection-Transformer
- Using Custom Datasets to train DETR for object detection
More references can be found in the project's final report
- Final Report Parper - PDF(TODO) or LinkedIn Post(TODO)