“Making Europe A Wilder Place!”

The project is a pioneering concept of creating wildlife reserves in diverse European landscapes. Rewilding is being implemented on a large scale throughout Europe. The use of rewilding principles, models, and tools is yielding tangible, observable, and long-term advantages for both nature and people. Collaborating with numerous partners, we are shaping a Europe that is more abundant in natural diversity and better equipped to handle climate change.

I got the opportunity to work with a small subset of their dataset. And it was still huge ! 

It had more than >378,000 cameratrap images captured. 

Due to the nature of cameratraps a lot of false positives are captured due to background movements (like grass moving in the wind)

The dataset has been annotated with the species class by expert biologist, however the bounding boxes are not available on the provided data. The ReWilding dataset was downloaded locally. Script for the download function is provided here

Since the ReWilding Dataset is not a public dataset, some of the private data has been masked in the scripts. 

A local instance of the MegaDetector v5 was set up for annotation of the images. Please follow the instructions provided in the MDv5 github repo. 

The images were annotated with bounding boxes and then merged with the expert species-level annotations. A threshold of 0.8 was selected for filtering bboxes with high confidence. This is heuristics and some [Reference]

The bounding box annotation was then converted into COCO, YOLO and TF format:

• COCO – x1,y1, w, h
• YOLO – xc, yc, w, h
• TF – x1, y1, x2, y2

For YOLO  (xc, yc) referes to the center of the bounding box, with the (w,h) being the width and height of the bounding box. 

Similarly (x1,y1) and (x2,y2) are the top-left hand corner and bottom-right hand corners of the bboxes. 

The dataset was split into train, validation and test splits (80%-10%-10%)

The label map was created for the species class.