Fauna-ing Over Object Detection I'm not going to copy my Github's README.md file or anything onto this page. This was meant more as an intro to the project and an explanation of research I did. I really just want you to click on the github link if you want some indepth results and a guide to my code.
I did this project in conjuction with BoulderAI. It is an object detection model trained on wildlife images from Imagenet. I had about a week to pipe, train, and evaluate the model. With this tight time constraint I was able to get the model to detect animals with a fair amount of accuracy (precision of 83%), but was unable to get the model to identify animal species. With more training time I am confident the model would identify some species correctly (due to the BW nature of the photos and the size of some animals I'm not sure it could differentiate between say, a robin and a sparrow).
I used TensorFlow's Faster-RCNN for this project. It allowed me to detect objects much faster than I could have with trianing a model ground up. This allowed me to start detecting animals after ~5 hours of training. I kept trying to tweak the model and what I was training it on to get it to classify the different species faster, but I think it might have just needed more time to train. I'm not really sure and object detection / classification is definetly something I'm interested in pursuing more. I have a few ideas of things I'll want to work on eventually. I'll probably start looking into this project and why I couldn't get classifications more when I start to work on those.
This project was a great intro to object detetion and computer vision. I also learned a lot about AWS, Python's os package, scoping issues, and working with companies/customers under a tight time constraint. The project whetted my object detection appettite, and I look forward to some more projects coming in the future. So stay tuned and if you have any cool computer vision project ideas, or TensorFlow API questions, shoot them my way!