If you've only got a few classes it's faster to transfer learn a model from a few hundred examples, get that model to relabel your entire dataset and then fix any mistakes it's made. This is less than ten lines of python using fast ai.

If you've split your labelled data into folders then you can scan through several thumbnails at once to catch any misclassified pictures.

Rinse and repeat until the model is good enough.

> and then fix any mistakes it's made

This might be the most practical way to approach this, but it's not really a fundamental improvement since you still need to consider every image (also the correctly labeled ones).

True, although as the sibling suggested you can sort by confidence and weed out the edge cases. It's also quicker to go through mostly correct images as you need to make fewer (corrective) interactions/clicks. You can even view eg four thumbnails at once to spot outliers.

> you still need to consider every image

Not necessarily. The classifier will not only tell you whether it sees a rat, but also its confidence. So you can manually classify just the low confidence images, retrain, and repeat until there's none left or you're happy with its detection rate.

Running a computer vision consulting shop, that's our way of life. For almost a decade we've developed semi-automated means to make the problem saner, but we have to maintain a network of people and software for labeling images.

Once you have each image in its own file, not hard to rig up a way to preview them one at a time and press a number key corresponding to the desired category (1–4 in this case) which will bin the image in that category and bring up the next image.

if you can do one per second that's just over 6 hours.

Faster if you can do them in a batch. If you have individual video segments, just label the whole segment. There isn't going to be no prey in one frame and then suddenly prey materializes from nowhere in the next frame.

Well, genius is 1% inspiration and 99% perspiration, in the immortal words of Thomas Alva Edison.

Yeah but it makes you wonder: is it genius, or is it just brute force pattern matching?

And what if he gets a new cat, or his cat switches to a different kind of prey?

From the outside, it might seem as if he is enslaved by his pet.

> From the outside, it might seem as if he is enslaved by his pet.

I don't see the issue, this is the explicit contract of having a cat for a pet.

> explicit contract

I think many first-time cat slaves would disagree

They'll learn.

It took me all of maybe 20 minutes

Brute force pattern matching would be cataloguing all possible matching pixel patterns. The whole point of the ML approach is that with a comparatively few images (yes even 20k is comparatively few) the computer will correctly match pixel-sets it has never encountered before.

If your comparison function returns a float instead of a binary value, then you don't need to match against every possible image. Instead you just return how much each image is "similar" to the given image. A simple approach is to use cross correlation for the similarity.