The book covers the whole job with images: classification, detection, segmentation, model architectures, dataset preparation, training, serving and monitoring. Of the copy I have I worked through the data and serving parts, which are the ones that concern me.
What I took from it
The dataset chapter changed how I answer these requests. The label is the expensive part: somebody has to say, image by image, what is in it. And before that somebody has to define the criterion, because if two people label the same photo differently, no model resolves the disagreement; it learns it.
The second idea is that capture conditions weigh more than the model. Light, angle, distance, background. A document photographed crooked on a cluttered desk is not fixed by a bigger model, it is fixed by a guide on how to take the photo. That applies directly to any project digitising receipts or invoices, which is a fairly common request where I work.
The third is the most commercially useful: nearly all of these projects that fail do not fail on the model, they fail because nobody budgeted the labelling or decided who does it. When I ask that in the first conversation, half the requests turn into something else, smaller and feasible.
What I left out
The model architecture chapters, a tour of recent research that ages quickly. And I will not opine on what I do not have: the copy is partial. Nor do I pretend that reading this qualifies anyone to run a vision project. It qualified me for the opposite, which is recognising when to hand the work to somebody who does this for a living.