DeTr and YOLOv6 for Object Detection using ExDark dataset

Project Description

This project aims to compare and evaluate the performance of transformer-based and traditional deep-learning object detection models on different image enhancement techniques.

ExDark Dataset

The Exclusively Dark (ExDark) dataset contains the largest collection of natural low-light images taken in visible light to date, including object level annotation.

Dataset Folders Structure

In git's repository root folder:

• ./ExDark/ExDark - Original images from ExDark's git repository, subfolers for categories
• ./ExDark_Annno/ExDark_Annno - Original annotations from ExDark's git repository, subfolders for categories
• ./ExDark_All - All images and annotations without subfolders
• ./ExDark_COCO - .JSON files for COCO format dataset generator (used by DETR)
• ./ExDark_YOLO(#TODO - Link Missing) - File for YOLO model training

Split

• 3000 images for training - 250 per class
• 1800 images for validation - 150 per class
• 2563 images for testing - rest of the images per class

Authors

• Ignacio Gomez Valverde (A20552714)
• Prashanth V.R. (A20531508)

References

More references can be found in the project's final report

YOLO

• You Only Look Once: Unified, Real-Time Object Detection

Transformer

• End-to-End Detection with Transformers
• Detection-Transformer
• Using Custom Datasets to train DETR for object detection

Anexes

• One Drive with the Dataset folder structure and Trained Models
• Project Intermediate Report - Word Document
• Project Presentation - Google PPT Document
• Project Final Report - Latex Document