Summary

The project was to detect each class (defined by us) on a magazine, then run a OCR on each specific box for each class and create a structured JSON containing every information across the entire magazine. It was a whole project through data annotation to delivery a Docker file, with both a web interface and a CLI API.

Skills Developed

Main Challenges

The main challenge with this project was determining when a article started and ended, we had to make several adjustments using digital image processing. Additionally, handling the variability in magazine layouts and ensuring accurate text extraction from different content types presented significant technical hurdles.

Summary

In this project I had to use a detection model (many models tested -RTMDet, FasterRCNN, RetinaNET, SSD - but the best case on the performance x processing time was YOLOm) but the main issue wasn't detecting the animals, for the project I had to count them, so I had to implement a tracker class, that would identify (i.e give an ID) each of the bbox returned from the model and count the number of IDs at the end.

Skills Developed

Main Challenges

The main challenges included:

• Training multiple models on our custom dataset
• Using the model to extract bounding boxes
• Identifying those bounding boxes using digital image processing techniques
• Implementing a robust tracking algorithm that could handle occlusions and similar-looking animals
• Ensuring accurate counting despite environmental factors in thermal imagery