3 ms·

Cars should not detect "obstacles". They should detect roads. They should not seek _exceptions_ to driveable space through classification- or segmentation-first

Cars should not detect "obstacles". They should detect roads. They should not seek _exceptions_ to driveable space through classification- or segmentation-first strategies. They should only detect driveable space. Full stop. You either detect a full-stopping-distance's worth of clear road through reliable (non visual) ranging with dense enough samples to preclude wheel-catching voids ... or you will continue to run into "objects" which slipped through your classifier. I can accept reduction in this margin (e.g. in dense traffic flow) when it is first shown to work with margin.

krainboltgreene 6y ago
> They should only detect driveable space. I'm interested to hear how you decide what's drivable without detecting things that decide if roads are undrivable.
- jvanderbot 6y ago
  You don't need to know it's a "car" or a "cow" or a "corvette" to know that the flat surface you are on ends in less than your stopping distance. Semantically, it's a meaningless distinction. But in implementation it's much easier to detect the shape of the upcoming terrain using, say, lidar or regular old multi-camera disparity. Flat and following your map: that's driveable. I'm guessing many take this approach, but it's seeming more likely that Tesla does not, and instead trusts monocular or arrayed cameras to track objects though frames from detection-first regimes. Seems dangerous. I'm glad you were interested.
  - dmitriid 6y ago
    A one-way street is also a drivable flat surface. As is a road with no lane markings. Can you guess why blindly driving down each of these will be dangerous?
    - toxik 6y ago
      Right, but can you guess why stopping in the middle of traffic for a fake stop sign is bad? Walk before you run is the essence of the argument here.
      - dmitriid 6y ago
        An automatic car which can only understand flat surfaces as drivable is not even walking. It's standing still, and is significantly more dangerous than a car slowing down and stopping in the middle of traffic. Realising that a flat surface is drivable is probably 1% of 1% of what an autonomous car does on a road. Same goes for humans, I'd say.
      - hashtagmarkup 6y ago
        > Walk before you run is the essence of the argument here. Walking on the highway because you got spooked is the essence of the reason why you're dead.
    - jvanderbot 6y ago
      You don't need only flat surface detection, but it should be the highest priority safety check. And it should be the primary check. We shouldn't turn onto one-way streets, of course! Just like we shouldn't swerve off streets when we detect something that can be spoofed ... such as the false positives shown in the article and the false negatives given by other Tesla crashes. You can actually correct those instances by using reliable ranging devices with dense enough sampling to ensure flat surfaces without intermediate obstacles. To say that there are no other factors at play is foolish, of course, but it's also foolish to say that somehow we shouldn't be doing this.
- hashtagmarkup 6y ago
  Obviously, roads should have minefields and electrified fences to ensure drivability. Potholes will be outlawed. Weather will be punishable by death.
  - jvanderbot 6y ago
    You can detect geometric hazards like potholes with reliable ranging devices. I'm sure what I said was poorly worded. What I meant was that reliable ranging with dense sampling should be the primary mechanism that self-driving vehicles use to determine what areas of the road are safe to drive on, and probably what is even a road at all. I think efforts to avoid this are based on cost, not safety or functionality. Other signals are just too transient to be safe. Vision-based classification-first regimes are not guaranteed to detect hazards as reliably as a point ranging device, is that controversial? Potholes exist, but are geometric negative space. Clues to their existence can be provided by imagery or other sensors (which are subject to all the weather effects also, by the way), but a positive detection of good road or pothole should have priority.
hashtagmarkup 6y ago
How did your country eliminate all the hoofed ruminants? Why did they do it? Do you have a death wish?
__m 6y ago
This wouldn’t work well with moving obstacles. The road could be clear until a person enters the road within your full-stopping distance.
- jvanderbot 6y ago
  Ok, I was too strongly worded, and you can read it as a reduction to basic ranging. That's fine. I can rephrase: Reliable ranging should override other sensor mechanisms for detecting driveable space and presence of hazards. Tracking with dense ranging (e.g., LIDAR) is not as challenging as the alternative: ensuring that non-ranging system will reliably segment all possible hazards.

krainboltgreene 6y ago

> They should only detect driveable space. I'm interested to hear how you decide what's drivable without detecting things that decide if roads are undrivable.

jvanderbot 6y ago
You don't need to know it's a "car" or a "cow" or a "corvette" to know that the flat surface you are on ends in less than your stopping distance. Semantically, it's a meaningless distinction. But in implementation it's much easier to detect the shape of the upcoming terrain using, say, lidar or regular old multi-camera disparity. Flat and following your map: that's driveable. I'm guessing many take this approach, but it's seeming more likely that Tesla does not, and instead trusts monocular or arrayed cameras to track objects though frames from detection-first regimes. Seems dangerous. I'm glad you were interested.
- dmitriid 6y ago
  A one-way street is also a drivable flat surface. As is a road with no lane markings. Can you guess why blindly driving down each of these will be dangerous?
  - toxik 6y ago
    Right, but can you guess why stopping in the middle of traffic for a fake stop sign is bad? Walk before you run is the essence of the argument here.
    - dmitriid 6y ago
      An automatic car which can only understand flat surfaces as drivable is not even walking. It's standing still, and is significantly more dangerous than a car slowing down and stopping in the middle of traffic. Realising that a flat surface is drivable is probably 1% of 1% of what an autonomous car does on a road. Same goes for humans, I'd say.
    - hashtagmarkup 6y ago
      > Walk before you run is the essence of the argument here. Walking on the highway because you got spooked is the essence of the reason why you're dead.
  - jvanderbot 6y ago
    You don't need only flat surface detection, but it should be the highest priority safety check. And it should be the primary check. We shouldn't turn onto one-way streets, of course! Just like we shouldn't swerve off streets when we detect something that can be spoofed ... such as the false positives shown in the article and the false negatives given by other Tesla crashes. You can actually correct those instances by using reliable ranging devices with dense enough sampling to ensure flat surfaces without intermediate obstacles. To say that there are no other factors at play is foolish, of course, but it's also foolish to say that somehow we shouldn't be doing this.
hashtagmarkup 6y ago
Obviously, roads should have minefields and electrified fences to ensure drivability. Potholes will be outlawed. Weather will be punishable by death.
- jvanderbot 6y ago
  You can detect geometric hazards like potholes with reliable ranging devices. I'm sure what I said was poorly worded. What I meant was that reliable ranging with dense sampling should be the primary mechanism that self-driving vehicles use to determine what areas of the road are safe to drive on, and probably what is even a road at all. I think efforts to avoid this are based on cost, not safety or functionality. Other signals are just too transient to be safe. Vision-based classification-first regimes are not guaranteed to detect hazards as reliably as a point ranging device, is that controversial? Potholes exist, but are geometric negative space. Clues to their existence can be provided by imagery or other sensors (which are subject to all the weather effects also, by the way), but a positive detection of good road or pothole should have priority.

hashtagmarkup 6y ago

How did your country eliminate all the hoofed ruminants? Why did they do it? Do you have a death wish?

__m 6y ago

This wouldn’t work well with moving obstacles. The road could be clear until a person enters the road within your full-stopping distance.

jvanderbot 6y ago
Ok, I was too strongly worded, and you can read it as a reduction to basic ranging. That's fine. I can rephrase: Reliable ranging should override other sensor mechanisms for detecting driveable space and presence of hazards. Tracking with dense ranging (e.g., LIDAR) is not as challenging as the alternative: ensuring that non-ranging system will reliably segment all possible hazards.