Detic - Detecting Twenty-thousand Classes using Image-level Supervision

This is cool! And interesting to see that a Fast-RCNN head is still being used in this day and age. As a learning exercise, I recently tried to replicate the Faster-RCNN paper from scratch. I got pretty close but had to peak at some existing code as the paper isn't quite clear on a few key details. I now have clean and readable (IMHO) implementations in PyTorch and TensorFlow 2/Keras that I'm writing some documentation for and will upload to github in the next few days.

What are the best object detector models these days? Is this it?

The last object detectors I played with were SSDMobileNet and Yolo. Their off-the-shelf performance was really bad for a cameras I set up pointing at my house and street. But this was years ago.

Does it differentiate pooping dog from not pooping dog? I’ve got a stubborn little critter that stealth poops, i could use another set of eyes.