Use Megadetector

[gsganden]

Megadetector is an object detection model for animals in camera trap images. We could use it to infer that an image is empty if there is no bounding box and to train on predict on crops rather than full images. The top entries in https://www.kaggle.com/c/iwildcam-2020-fgvc7 all used some version of this approach.

[gsganden]

Megadetector has separate "animal" and "human" categories, so we could use it out of the box to classify images as empty, containing humans, or containing animals.

[shardulparab97]

@gsganden have loved the idea and potential impact of AutoFocus and would love to contribute to the same. Is it fine if I pick up this issue?

[gsganden]

Go for it!

I did a first test using Megadetector to classify images as containing humans or not and as containing animals or not; those results look good. I am planning to follow up with more and bigger images. You are welcome to try out that approach too if you would like, but there would presumably be some redundancy between your work and mine.

I do not know when I will have time to train a classifier for specific animal types on Megadetector-generated crops. I would love for you to try that approach if you are up for it.