Detecting stacking of products

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

I am trying to build an object recognition system for a retail store. I have built a network using Yolo algorithm and Resnet backbone and I can successfully identify products in a frame. However, I want to also identify and count products when they are stacked either vertically or when placed behind one another.

image
image1810×1180 1.14 MB
image
image1912×1311 1.25 MB
image
image1958×1837 1.32 MB
image
image2048×1717 1.41 MB
image
image2372×1236 922 KB

So as we can see in the images, the system detects kettle chips and too yumm chips when placed together, but two packets of too yumm chips/buttermilk/coke are identified as a single product. How can I differentiate these as two separate products along the edge? Basically I want the model to output two separate bounding boxes even when they are placed close to each other.

2

It depends on more training dataset which contains chips/buttermilk/coke close to each other.
Adding these images and labels, then trigger training.

3

Do you think training using Mask RCNN would help in this problem?

4

If you want to use Maskrcnn, please try to use faster-rcnn to train at first to see what is its result.

More, actually try a bigger backone with yolo, for example, if you were using resnet10, then just try to use resnet18.

5

Sure, thank you. Will try.