When memory was measured in kilobytes: The art of efficient vision | Hacker News Reader

When memory was measured in kilobytes: The art of efficient vision

softwareheritage.org

155 pointstodsacerdoti34 comments

One approach that blew my mind was the use of FFT to recognize objects.

FFT has this property that object orientation or location doesn't matter. As long as you have the signature of an object, you can recognize it anywhere!

I believe orientation still matters but you’re right that position doesn’t.

FFT is equivalent to convolution, which is widely used today for object recognition in CNNs.

> FFT is equivalent to convolution

What do you mean by that? Could you give me an example?

The FFT, composed with pointwise multiplication, composed with the inverse FFT, is equivalent to convolution. The FFT is not.

The basic convolution theorem.

https://en.wikipedia.org/wiki/Convolution_theorem

That is something else entirely.

Then if you know what the OP meant why did you ask?

because they made a nonsensical claim that doesn't align with my (and likely their) understanding of what the FT is and does.

The FT is _NOT_ just a convolution, but under certain conditions a specific operation on FT terms is equivalent to a convolution.

I didn't know what they meant. There are so many FFT tricks. I was hoping this was another.

I want to believe that however obsolete these old algorithms are today, at least some aspects of the underlying code and/or logic should prove useful to LLMs as they try to generate modern code.

The idea that ML is the only way to do computer vision is a myth.

Yes, it may not make sense to use classical algorithms to try to recognize a cat in a photo.

But there are often virtual or synthetic images which are produced by other means or sensors for which classical algorithms are applicable and efficient.

I worked (as an intern) on autonomous vehicles at Daimler in 1991. My main project was the vision system, running on a network of transputer nodes programmed in Occam.

The core of the approach was “find prominent horizontal lines, which exhibit symmetry about a vertical axis, and frame-to-frame consistency”.

Finding horizontal lines was done by computing variances in value. Finding symmetry about a vertical axis was relatively easy. Ultimately, a Kalman filter worked best for frame-to-frame tracking. (We processed video in around 120x90 output from variance algorithm, which ran on a PAL video stream.)

There’s probably more computing power on a $10 ESP32 now, but I really enjoyed the experience and challenge.

This was our vehicle: https://mercedes-benz-publicarchive.com/marsClassic/en/insta...

That's awesome! What kind of hardware was needed to pull that off? And was the size of the bus any indication of the answer?

You could even argue that ML does classical vision in addition to other stuff.

CNNs learn gabor filters. The AlexNet paper even shows this [0]

Or if you look at the work ViT built itself on, they show attention heads will also learn these fillers. [1] That's actually a big part of how ViTs work, the heads integrate this type of information

[0] https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3...

[1] https://arxiv.org/abs/1911.03584

Any recommendations on background reading for classical CV for radar?

I don’t know anything about radar. I have a book called “machine vision” (Shmuck, Jain, Kasturi) easy undergrad level, but also very useful. It’s $6 on Amazon.

Kasturi was my undergraduate honors advisor!

Small world! These are always just names on a book to me.

Awesome, thanks!

Don't know about radar but here's a good book on classical CV https://udlbook.github.io/cvbook/

even though I think Simon admits that most of it is obsolete after DL computer vision came about

> is obsolete after DL computer vision came about

I just don’t understand this. Why would new technology invalidate real understanding and useful computer algorithms?

Maybe… some of these algorithms from the 1980s struggled to do basic OCR, so they may need a lot of modification to be useful.

That whole approach of "find edges, convert to line drawing, process a line drawing" in the 1980s struggled to do anything at all.

There was a surprising amount of useful OCR happening in the 70’s.

High error rates and significant manual rescanning can be acceptable in some applications, as long as there’s no better alternative.

I find that modern OCR, audio transcription, etc... are beginning to have the opposite problem: they are too smart.

It means that they make a lot fewer mistakes, but when they do, it can be subtle. For example, if the text is "the bat escaped by the window", a dumb OCR can write "dat" instead of "bat". When you read the resulting text, you notice it and using outside clues, recover the original word. An smart OCR will notice that "dat" isn't a word and can change it for "cat", and indeed "the cat escaped by the window" is a perfectly good sentence, unfortunately, it is wrong and confusing.

Thankfully, most speech misrecognition events are still obvious. I have seen this in OCR and, as you say, it is bad. There are enough mistakes in the sources; let us not compound them.

I'm not sure I can sign on to this. In particular, this sounds kind of like an indictment of many algorithms. But, how many where there? And did any go on to give good results?

Considers, OCR was a very new field, such that a lot of the struggle was getting data into a place you could even try recognition against it. It should be no surprise that they were not able to succeed that often. It would be more surprising if they had a lot of different algorithms.

Amazing. Wonder how fast it would be on a modern computer

+1, which is faster when compare to OpenCV findContours

I still deal with <128kb ram system everyday

Ah, Mac user?

more like STMicroelectronics user

I too need a coffee in the morning before I can do anyhting

Reply on news.ycombinator.com