Machine Learning Development
A walkthrough of the ML development process for wildlife detection.
1. Objectives
Machine Learning Development will examine model selection, model training and deployment on the Raspberry Pi.
- To innovate on annotation - It takes a long time to annotate: 2 weeks for human annotation compared to 2 hours for the model.
- Implementing new technology is exciting
ML Objectives
2. Conventional Classification
We started with conventional animal classification using MobileNetV2 (for this we selected 1146 images which had only one species of animal).
Impala Classification
Waterbuck Classification on Uncropped Background
Waterbuck Classification on Cropped Background
Monkey Classification on Uncropped Background
Monkey Classification on Cropped Background
For images with multiple species, classification isn't the best resort. Also, cropping out the background had a significant positive effect.
Snippet of multiple animals per image examples
3. Animal Detection
To handle images with multiple species, we moved to object detection approaches.
Animal Identification Example
Impala Identification Example
Impala Video Identification Example