AI for Bears
Classifying and identifying bears on low-powered edge-hardware
Premium participation: Get 1-on-1 Mentoring & Guidance
Try a free mentoring sessionBears, initially perceived as cute, become "problem bears" when they enter villages to attack livestock. Once a location with food is found the bear will always return. This behavior can lead to dangerous situations, often resulting in the killing or relocation of bears. To address this, we aim to develop a real-time bear-detector system on low-power hardware to activate repellents and prevent such conflicts.
Technical Aspects of the Challenge
Bear Presence Classification: Create a small convolutional neural network (CNN) that accurately identifies the presence of a bear on edge-hardware. Models like VGG, MobileNet, ResNet, EfficientNetB0 are recommended.
Object Detection Algorithm: Develop an algorithm that detects bear faces on edge hardware, preferably Linux-based systems with NPUs. YOLO v5/v8 is a promising algorithm for this task.
Identification Model: Design an identification model to recognize individual bears and distinguish new ones. Consider Siamese CovNext with triplet or circle loss, Light Glue, LoFTR as promising algorithms.
Real-World Impact of the MVP
Our goal is to reduce human-bear conflicts in the five communes of Romania by 50%, achieved through the implementation of the developed MVP.
Quick Facts
Bears look cute but can pose threats. Public perception of bears is changing, with increasing fear in Romania. Human-wildlife conflicts lead to a rising number of bear casualties.
Data Description
We provide datasets with bear images, including: BearID: 4675 bear face chips, 132 bears represented, and 1500 images with face detection annotations. Hack the Planet: 50k photos, 20k containing bears with MD-generated bounding boxes. And more…
Info session
Our challenge partner
sdcvsdvs