Hacker News Reader

Hacker News

Comment by ramshanker | Hacker News Reader

Full thread

ramshanker

Follow up question I googled.

What is the resolution of our Eyes?

Ans: 576 Megapixels [1]..... Holy..... And here we are, trying to rely on couple of cameras for autopilot.

[1]http://www.clarkvision.com/articles/eye-resolution.html

View on HN

The resolution of the camera has nothing to do with our ability to understand the road. I can drive reasonably well with just a camera feed of the road, but the computer can't. So the bottleneck is in the image processing capability of the machine, not the quality of the camera.

Correct. You could follow a 1970's Atari green/black pixelated set of lines just as well. And optical illusions such as the Kanizsa Triangle demonstrate even that much info is more than ample, since your brain only needs the tiniest of hints to calculate shape and size in order to calculate direction.

https://www.illusionsindex.org/i/kanizsa-triangle

Your comparison to a simple driving game might be misleading. A two-color game is probably an easier challenge, not a harder one, because the game developer has reduced your input to almost pure information rather than throwing a lot of visual data at you to interpret.

In other words, it isn't that your eyes and brain are so good that you can still drive at an Atari level--it's that the Atari level is optimized for your eyes and brain.

I agree to an extent, but it is a complex comparison to make, so it can easily look that way.

Our brain creates a model of the surroundings based solely on memorized bits of the environment (allowing us to blink without the world going dark, and to track things in our peripheral views) plus prediction, which is something Atari exploits so we can convince ourselves the bars and dot have motion. We even temporarily convince ourselves that the blooping is the sound of impact.

A long way to say that our sensors don't send complete objects like an OOP function would. Instead, they transmit just enough info (hints, maybe) to refresh the mental model. Similar to how we can see 4 dots and mentally translate it into a square laying in some particular direction in 3D space, even without lines or actually being 3D.

The "bit" analogy breaks down when you consider this aspect. The data our brain works with isn't the same as the data a computer works with.

Yes but keep in mind that this data is immediately and heavily compressed and encoded - the data sent into the brain from the retinal ganglion cells has been estimated to be on the order of 10s Mbit/sec from careful studies on rodent RGN statistics. https://doi.org/10.1016/j.cub.2006.05.056

Not an expert, but this is also my understanding: The nerves are more like a neural network, not just wiring to the brain: https://heartbeat.fritz.ai/the-ancient-secrets-of-computer-v...

It’s not just resolution; the dynamic range (the “f-stops”) of the human visual system is incredibly wider than even high-quality film and digital cameras, meaning that cameras are much less capable at handling details between wide ranges of light and dark.

Though post-processing from neural nets seems to close this gap.

https://www.youtube.com/watch?v=bcZFQ3f26pA

> Ans: 576 Megapixels

I’m skeptical. Since the human eye has less than 150M sensor cells, and our sensors aren’t as good as 8 or 10 bit resolution pixels, that answer is overestimating by at least 4x, and possibly a lot more. https://en.m.wikipedia.org/wiki/Photoreceptor_cell

He misconveyed the source. The claim was about the detail the eye can percieve in a given field of view, over time and by moving the eyes over it, not the amount of information the eye captures at any one time, which is vastly lower.

It's mostly irrelevant for cars TBH, since crashes are quick events.

That’s a simplification. The events leading up to a crash are important. The spatial understanding of the situation on the road, the behavior of other drivers and other factors all play a role in the seconds before a crash.

Yes, but there's a time-accuracy trade off when you see things by scanning around. A good, wide camera trades off remembered perepheral details for lower-resolution but temporally up-to-date details, which matters more in a crash I'd expect.

I can tell the difference between two 24-bit colors that are 1 value off in 1 channel, and that's not counting that digital displays can't fully exploit the darkest or brightest ranges of our eyes' sensitivity. Have you ever seen an HDR display showing well-made HDR content? Spatial resolution is not so easy, since we have much higher resolution at the center of our FOV.

Sure, that is true, you can sometimes see the difference in two 8-bits-per-channel colors. But, it takes both time and more than one pair of photoreceptive cell in your eyes to see that difference. There’s a tradeoff between color resolution and spatial resolution. Your spatial resolution at the sensitivity of seeing the difference in 8-bit colors is way lower that 150 megapixels.

BTW, 24 bit colors that are carefully done, paying proper attention to color space and gamma, have been specifically designed to be at the JND (just-noticeable difference) with a 1-bit change in value. If you can always and reliably see the difference between any pair of 24 bit colors that are only 1 bit different, then you have abnormally good color vision. It’s statistically unlikely for that to be true, but some people do have it.

> Ans: 576 Megapixels [1]..... Holy..... And here we are, trying to rely on couple of cameras for autopilot.

We only have high resolution at the point of gaze point. Everything else is blur, so it's not all comparable with a camera.

Worse: we think we have high resolution outside of that point, but it's actually our brain making up the details it thinks are there. We don't even really have colour outside a narrow field, and all of that is obfuscated by blood vessels dropped in front of core parts of our visual field.

Resolution has very little to do with object detection and obstacle avoidance.

(not saying Musk isn't full of shit)

Reply on news.ycombinator.com