I will have the opportunity to participate in biology field research in Peru, so I thought I'd come up with a solution to a relevant problem there related to ML. I learned that researchers spend months labeling camera trap data from wildlife reserves before they can perform analysis on it. This means that they will be at least a few seasons late to respond to any changes in animal populations, biodiversity, etc. that may be caused by spontaneous changes in the climate.

What it does

With data that I previously got from an Operation Wallacea Reserve in Peru, I trained a Faster-RCNN object detection model fine tuned with the iNaturalist Faster-RCNN on 32 classes of animals that are commonly seen in the Peruvian Amazon, achieving a high degree of accuracy on prevalent species.

How I built it

I learned the Tensorflow Object Detection api from scratch and implemented the infrastructure to train my specific model. The data model was trained on a paperspace ML-in-a-box machine over approx. 10k steps, achieving a final average loss of < 0.5 (Though I have a strong feeling that it overfit)

Accomplishments that I'm proud of

The model actually works

What I learned

-tensorflow is very finnicky and there is not a truly straightforward plug-and-play object detection model training method. -How data is setup for model training. -linux is harder to use than windows but much less vague.

What's next for The Project

In the future, perfected versions of this model could be incorporated into scalable software that would improve over time and improve quality of life for biologists conducting field research.

Built With

  • jupyter-notebooks
  • object-detection-api
  • paperspace
  • python
  • tensorflow
Share this project: