Telling the time with computer vision | Hacker News Reader

Telling the time with computer vision

blog.jinay.dev

47 pointsjinay17 comments

A fun project.

Separately: it's a step backwards. The advantage of the analogue clock is that you can get a very rapid idea of how much time you have left, or if something is a half hour away etc without having to parse a higher-resolution answer and then do math. In fact for the case at hand (doing an exam) it's exactly what you want.

None of which is criticism of the project of course!

That works really well if you're doing a task that ends at on the hour.

But if your task runs from 10:47 - 13:32 then looking at an analogue clock and seeing how far through you are is at least as tricky as it is with a digital one.

Sure but that's an extremely uncommon case. If anything scheduled events, if not set to hour/half/quarter are set on 5 minute boundaries.

I can see where you're coming from. In my experience, people who can read an analog clock well can do it faster than someone reading a digital one.

Exactly. And all that computation isn't something you want to do in an exam; you just want to glance up and realise there's no need to panic. Of that you're half way through the time but not half way through the questions :-(. Either way a single-bit question.

I appreciate that the author was looking to solve this problem using a _classical_ CV approach just like in the good ol‘ days as opposed to just train another ANN that predicts the time for a sufficiently large training set.

As for the detection of the hands itself: I think I‘d try to look for them in fourier space to get a more robust result (to address noisy dials like in the last examples).

Thanks for the advice! Could you elaborate a bit more on what you mean by searching in fourier space? Not too familiar with how that works.

I was hoping for a one-shot learning approach.

>Few things are more embarrassing than having to spend 20 seconds trying to figure out how much time you have left on a standardized test

This is a sorta-clever technique ... but if you can't have your phone, this is 100% useless

There are only three proper ways to worry about how much time you have left on a test:

- don't worry: keep going until you're done or you run out of time (the best option)

- wear a watch

- periodically look at the countdown timer on the test (if it's either being done on a computer, or if there's a timer in the room)

You're completely right. If you had your phone you could just look at the phone time anyways :)

As I mentioned in the conclusion, there wasn't really meant to be any practical use for this project.

> there wasn't really meant to be any practical use for this project.

I immediately thought of a (sort of) practical use for the project, syncing the time on a 20th century electric master/slave clock system. If you don't know, slave clocks ran off a master pendulum clock which sent them an electric pulse once every minute to advance their hands and keep them all in perfect sync. I built a master clock replacement out of an ESP32 dev board that keeps all my slave clocks correct using NTP, but I still have to manually adjust if they get out of sync due to a power outage. If the master clock could see them and read the dials, it could sync them without any intervention.

Additionally, I've got an idea for a automatic time setting system for an 18th century grandfather clock that would be a lot easier if it could visually read the hands.

I could see a potential for usefulness if you wanted to categorize all of the images you had of clocks with what times are listed to be able to put them in [quasi] chronological order

And then - maybe - auto-select a clock image that matched the publish time, for example, of a blog post?

there was a website that did this years ago. Manually selected, I presume, but any time of day it had a different clock with the correct time shown :)

edit:

and then there's this: https://humanclock.com/

You know what I call a project without a practical use? Art.

It's certainly an amusing rabbit trail, though :)

> don't worry: keep going until you're done or you run out of time (the best option)

This really depends on the type of test. For many tests, checking your pace and adjusting how long you spend on each question based on remaining time is an important strategy. For many tests, you would be much better off with a less perfect answer on all the questions than one really good answer on one question and no answer for all the others.

Oftentimes, you should use the total amount of time to determine the level of detail and effort the test givers are expecting you to use for each question.

Reply on news.ycombinator.com