4 ms·

Pretty cool. Why did they use an input that does NOT provide any information about depth/distance from objects?

sqeaky 9y ago
I can only guess because I am not on their team, but: They clearly didn't need it. Human pilots clearly don't need it. Extra custom sensors might produce more noise than they are worth. Image processor is a hot topic of research in CNNs.
- analyst74 9y ago
  Human pilots do get depth information by either looking, and/or by radar and other instruments.
  - skj 9y ago
    Humans don't get depth information past some distance that is probably pretty small compared to how far away objects are in flight. Past that point it's all contextual.
  - djrogers 9y ago
    Human pilots don't really use vision to see depth when flying - most things in flight are too difficult to judge distance/size, so one learns to fly without. As for radar, most planes don't have their own radar. About all you're gonna get for anything close to depth is al altimeter.
    - robertfw 9y ago
      to be fair, we're also usually not flying around inside a room
- otto_ortega 9y ago
  Except humans derive distance/depth information from their sight (depth perception), that's why we have a pair of eyes and not a single eye. The equivalent on this setting would be adding a second camera. I'm not criticizing, their experiment is pretty cool, I was just wondering why they chose to use only the camera on board.
  - blt 9y ago
    Humans can fly quadrotors very fast and precisely through monocular FPV cameras.
  - sqeaky 9y ago
    > that's why we have a pair of eyes and not a single eye. This is really overstated. It really only matters for about 3 or 4 meters of distance. We do depth perception well enough at distances to drive cars. We also do just fine at perceiving depth in video games and single lens camera. How often do you have trouble determining the depth of something in a movie? Only about as often as the filmmakers want you to.
alexleegk 9y ago
It's also probably because it's a lot easier to use an off-the-shelf commercial drone than to build/modify one (from a researcher's perspective), and commercial drones typically already have cameras than can be accessed through their SDKs. Also, off-the-shelf depth sensors can add a lot of weight to the drone. It might still be possible to fly with the extra weight, but now the drone will be more sluggish and fragile. It would be great if commercial drones had a built-in depth sensor. Distance sensors such as sonar and proximity sensors are usually very noisy and they are susceptible to interference (if you use more than one).
- otto_ortega 9y ago
  Those drones are perfectly able to lift a couple of pounds according to their specs. The other reason you mentioned about not wanting to customize it is more likely.
  - alexleegk 9y ago
    Yes, but what I was trying to get to is that even if the drone can handle the payload, the drone becomes too heavy to crash safely, and it might break more easily (after falling down with all that extra weight).
  - blt 9y ago
    Vision-based navigation is ultimately more interesting for robotics. Structured light depth cameras don't work in bright sunlight and aren't useful for long range perception. They also use more battery power for the light source.

I can only guess because I am not on their team, but: They clearly didn't need it. Human pilots clearly don't need it. Extra custom sensors might produce more noise than they are worth. Image processor is a hot topic of research in CNNs.

analyst74 9y ago
Human pilots do get depth information by either looking, and/or by radar and other instruments.
- skj 9y ago
  Humans don't get depth information past some distance that is probably pretty small compared to how far away objects are in flight. Past that point it's all contextual.
- djrogers 9y ago
  Human pilots don't really use vision to see depth when flying - most things in flight are too difficult to judge distance/size, so one learns to fly without. As for radar, most planes don't have their own radar. About all you're gonna get for anything close to depth is al altimeter.
  - robertfw 9y ago
    to be fair, we're also usually not flying around inside a room
otto_ortega 9y ago
Except humans derive distance/depth information from their sight (depth perception), that's why we have a pair of eyes and not a single eye. The equivalent on this setting would be adding a second camera. I'm not criticizing, their experiment is pretty cool, I was just wondering why they chose to use only the camera on board.
- blt 9y ago
  Humans can fly quadrotors very fast and precisely through monocular FPV cameras.
- sqeaky 9y ago
  > that's why we have a pair of eyes and not a single eye. This is really overstated. It really only matters for about 3 or 4 meters of distance. We do depth perception well enough at distances to drive cars. We also do just fine at perceiving depth in video games and single lens camera. How often do you have trouble determining the depth of something in a movie? Only about as often as the filmmakers want you to.

alexleegk 9y ago

It's also probably because it's a lot easier to use an off-the-shelf commercial drone than to build/modify one (from a researcher's perspective), and commercial drones typically already have cameras than can be accessed through their SDKs. Also, off-the-shelf depth sensors can add a lot of weight to the drone. It might still be possible to fly with the extra weight, but now the drone will be more sluggish and fragile. It would be great if commercial drones had a built-in depth sensor. Distance sensors such as sonar and proximity sensors are usually very noisy and they are susceptible to interference (if you use more than one).

otto_ortega 9y ago
Those drones are perfectly able to lift a couple of pounds according to their specs. The other reason you mentioned about not wanting to customize it is more likely.
- alexleegk 9y ago
  Yes, but what I was trying to get to is that even if the drone can handle the payload, the drone becomes too heavy to crash safely, and it might break more easily (after falling down with all that extra weight).
- blt 9y ago
  Vision-based navigation is ultimately more interesting for robotics. Structured light depth cameras don't work in bright sunlight and aren't useful for long range perception. They also use more battery power for the light source.