Google’s Object Detection API vs. Yolov3
This post is sponsored by SugarKubes, a container marketplace for developers. Check out some of our latest demos.
The common perception of Google, Amazon, Microsoft is that their AI systems are state of the art, and broadly speaking they are, or more accurately could be.
I’ve been working with Google’s object detection API recently in the hopes that if it was amazing, I would be able to use it to get highly accurate bounding boxes before it ever hit a human annotator. That way annotators are only adjusting bounding boxes and adding ones that were missed. I was wrong.
Feel free to demo the yolov3 api yourself here.
As you can see Google’s Object Detection API pretty much sucks compared to an off the shelf Yolo v3 model trained on coco data. Now it’s important not to draw the wrong conclusion. Even as I’m posting this they could have updated the model and now it’s detections are absurdly accurate under all kinds of conditions but the reality is that it would be difficult for Google to beat a small team focused on a vertical problem with vertical-specific assumptions built into the model and more importantly the product.
Train on!
