AI for Earth Camera Trap Stuff


We’ll talk about...

Our repo

All our stuff is available on GitHub. If you remember one thing from this presentation, make it this link:

    github.com/agentmorris/MegaDetector

Oh, and this:

    cameratraps@lila.science

Detectors and why they’re awesome

What’s a detector?

A classifier tells you what an image is:



This image is warthog.

– A classifier

A detector tells you where in the image something is:



There’s a thing here.

– A detector

Detectors can also identify things:



There’s a warthog here.

– A multi-class detector

Even though we don’t care about where in an image animals occur, and we do care about species classification, we want to convince you that it’s useful and important to build a “generic” detector for camera traps.

Benefits of generic detectors

Why bother with detectors?

  1. Natural way to get rid of non-animal images
  2. Get value from lots of data, even in new ecosystems
  3. Make classification a lot easier:
    Classifying an animal vs. classifying an animal and lots of background goop
  4. Naturally handle classifying multiple species in an image
  5. Naturally handle training data that’s labeled at the sequence level, and is unreliable at the image level
  6. Natural way to get rid of non-animal images (so important we’re saying it twice)
  7. Natural way to get rid of non-animal images (and a third time, now in bold)
  8. Natural path to counting (sort of)




Our detector... aka “MegaDetector”, pronounced “MegaDetectooooooor

  1. Available (code and model) on our GitHub page
  2. Trained on several hundred thousand bounding boxes from a variety of ecosystems
  3. Mostly about animals, but we recently added a “beta” people class, about to improve this and add a “vehicle” class


Workflow

  1. Batch API


  2. Results on real data!
    Live demo, post-hoc viewers will have to settle for a teaser image...


  3. Timelapse integration
    Live demo, post-hoc viewers will have to settle for a teaser image...

  4. Image credit RSPB, Timelapse credit Saul Greenberg

  5. Other ways to run our detector(s)...
    1. Handy scripts on our GitHub repo
    2. Download our model and run it however you like (email us or buy us fancy coffee please)

How classifiers fit in to our universe

We train per-project (i.e., per-ecosystem) classifiers on images that have been cropped by our (generic) detector. We use this two-stage approach instead of a whole-image classifier or a multi-class detector.

I.e., sadly, we believe we’ll never be able to build a “MegaClassifier”.

Fun demo of some classifiers we’ve released publicly(ish)? Yes please!

Live demo, post-hoc viewers will have to settle for a teaser image...




Data

We’ve worked with lots of our camera trap friends to make lots of labeled images available at:

http://lila.science

Go play with data, and send us new data sets!