0

For instance, if I was trying to detect (and segment) people in a "Where's Waldo" image (an image dense with hundred of people in each image) and label each person based on the color of the shirt they are wearing, would a small dataset (a total of 10-100 images for training and validation) be sufficient because there are so many instances of these "objects" in each image (assuming I use pre-trained weights from COCO)? Is it vital to have a large dataset (> 1000 images) for these cases (assuming you are using some detection-segmentation algorithm like Mask R-CNN)?

Another perspective from which you can look at the question:

What matters more, the number of images you train your segmentation algorithm on or the number of instance of detectable objects over the span of the entire dataset?

pavpanda
  • 66
  • 7

1 Answers1

0

That's a good question. If you pose this is a color based segmentation (plus clustering on top of that to detect "objects"), you might be able to get away with fewer images since the clusters ("people") are typically distinct.

I'd also recommend incorporating priors instead of just feeding data (since you have less data). How does color based segmentation using say Gaussian Mixture Models work? How do clustering techniques like Incremental GMMs work? If these don't work, you can try combining these with deep learning techniques. In general, less data = incorporating priors to get the job done!

Sai Krishnan
  • 116
  • 3
  • Had this not been a color-based segmentation problem, would you still suggest incorporating priors? – pavpanda Feb 10 '20 at 20:48
  • If possible, incorporating some priors is a great way to get a solution. If it's a research project and you're trying to arrive at a solution without incorporating priors, well, that's an entirely different ballgame. Also given that you have less data, incorporating priors is not a bad way. As an alternative, you can do data augmentation with the few (ish) data points you have and see if the network at least overfits to the training data. Then you can see how little data you need to train it. – Sai Krishnan Feb 11 '20 at 14:30