I don't see any subtitles only some innocuous captions in the blue box.
In any case, from what I see of this video, QCOM did a surprisingly decent job of depth sensing given the little time apparently they've been working on it.
Meanwhile Intel has gone from generation to generation and seems to me that QCOM has leapt ahead of Intel technically, performance wise, and with alot less time and money spent.
The depth map performance that QCOM has shown here is better than what I've seen with the various Intel Realsense platforms. And even better, you don't need pricy Intel SoCs to run the sensor. Just less pricey QCOM processors no doubt!
I am the owner of this video (and the talented actor:)).
I've removed the subtitles commentary after a good advice that it might be too hursh on Qualcomm...
I want to make this clear - I have only but the utmost respect to Qualcomm and its employees as I've worked with them very close in the past but some of their remarks after being asked about Realsense in comparison with their technology was just not true ("Qualcomm claims that according to its understanding, the thermal overhead on Intel RealSense is a barrier to entry for smartphones. That means it most likely runs too hot to work well in a handset").
also, the choice of the technology together with its late-entry shows its immature nature as seen in their video where the dark key are not detected (just imagine what would happen outdoor\sunlight conditions).
also, there was a true statement here that we have developed several generations and technologies at Realsense and I definitely see this as a big advantages real-world usages and HVM requires a lot of experience that only few companies poses and there is a HUGE gap between showing a nice (and environmentally controlled) video.
Hi Roi, What chipsets will this run on now that Intel seem to be exiting the mobile market with Atom etc? Will Intel be licensing out the platform to run on other ARM units?
low reflective objects are always an encounter to any sort of sensor that uses active illumination. the same is valid for conditions with a high ambient light level. the straight forward answer to both is "use more light". and the "why not?" reasons are ranging from eye safety concerns to power (battery and thermal) downsides.
its nice being able drive the performance to new limits but using the mentioned nearby options is just not acceptable. the box of options and thus most of the limitations is drawn by the intended application case. asking for breaking the box will render the sensor set invalid for the mentioned case. smart solutions like polarization based 3D scanners might be a big step ahead for some aspects without breaking the box - but even those concept will not be able to solve anything.
maybe you did not understand that a technically impossible item (according to the basic measurement principle: some light gets reflected from the object) is something that will never solve.
for comparing sensor systems it sounds more reasonable to me to look at the performance in the areas where the device should be functional by it's concept.
when it comes to technologies that relies on light coding\modulation than you are right.
don't forget that stereo is matching a right pixel to the matching left one so its much more robust when its comes to detecting low reflective or high illuminated targets
as for the host platform - we are providing an open-source driver that can be compiled to any host platform (Intel\ARM) and OS (Linux\Win\Android\OSX). as long as it has USB\MIPI
still stereo needs some visible structure on the surfaces. else it can only take edges for depth hints. so its going to be a rough estimator in the critical cases where the scenery is dark. especially curved objects will turn out getting get hard to catch correctly as they dont expose solid outlines but will steadily change shape for different view points.
with a pencil and a piece of paper you should be quickly able to verify whats the difference in detection between a black disc and a black sphere. i would say the sphere will be seen a bit "wider" but on the right depth. the second will not be true for real world curved objects as they can have any sort of curvature not only perfect circles. edgy object instead have much of a chance exposing faces in a very high depth ranges where one camera can see much of it and the other will only see few to nothing. so if you have not enough structure on the surface for detection then detection will suffer and thus the 3-dimensional shape of your scene objects does matter quite much and thus will teach you the (un-avoidable?) precision and detection limits of your device. Anyone operating a 3D sensing device sooner or later came to the conclusion that a quality meter is needed providing areas amongst others with indication of low trust to even no detection. that happy you might be on what is detected - dont forget you will have areas that failed detection. should you worry about the "unknown"? in safety and security applications you definitely should. a system for such a case will have the chance to cease working any now and then whilst using a safe-state strategy that might be sort of a shutdown. the more real world you go the more you will discover your non-detection cases. you will have to put it via the filter of "does it matter to my systems functionality?" - for a device that is compareable to a web cam service level it wont matter in most cases. but you probably dont want to stop there...
the comments are focused to Qualcomm's own performance claims:
1. number of depth points - they can do 10k vs RS400 which can do 1M 2. depth precision is 0.1mm vs 0.01mm (both are best case I assume) 3. working environment - we are able to work at any ambient environment (Dark->direct sunlight) where SL would have great difficulties. just look at the keyboards of the Qualcomm video and notice that the dark keys are not detectable due to their low reflection. 4. Qualcomm claims the RS is not suited for mobile devices while we are probably the perfect match due to our low form-factor\power and robustness.
Depth resolutio of both these solutions won't certainly be 10 or 100 mic's. 10 microns is a ridiculous claim. While I haven't seen the RS400 yet, I'd bet a good amount of money it's not even sub 1 mm in depth resolution. Also the claim about direct sunlight performance is ridiculous too. And not just because the previous Realsense generations were useless with any sunlight cast on the scene.
If real life precision is 10 times higher than your theoritical claim, this would make 100 microns precision which would still be very precise for such a camera. Or even 100 times worse is still very good performance.
I could not find anything about RS460, is that different than D410 or D435 which are going to be released soon?
You mean you need to have money at risk before proving your claim? I'd say you make the claim, you should provide some data to back it up. Otherwise...
I don't see any subtitles only some innocuous captions in the blue box.
ReplyDeleteIn any case, from what I see of this video, QCOM did a surprisingly decent job of depth sensing given the little time apparently they've been working on it.
Meanwhile Intel has gone from generation to generation and seems to me that QCOM has leapt ahead of Intel technically, performance wise, and with alot less time and money spent.
The depth map performance that QCOM has shown here is better than what I've seen with the various Intel Realsense platforms. And even better, you don't need pricy Intel SoCs to run the sensor. Just less pricey QCOM processors no doubt!
I'd score it QCOM 1, Intel 0 so far.
Yes, indeed, the owner of this video has removed the subtitles.
DeleteHi,
ReplyDeleteI am the owner of this video (and the talented actor:)).
I've removed the subtitles commentary after a good advice that it might be too hursh on Qualcomm...
I want to make this clear - I have only but the utmost respect to Qualcomm and its employees as I've worked with them very close in the past but some of their remarks after being asked about Realsense in comparison with their technology was just not true ("Qualcomm claims that according to its understanding, the thermal overhead on Intel RealSense is a barrier to entry for smartphones. That means it most likely runs too hot to work well in a handset").
also, the choice of the technology together with its late-entry shows its immature nature as seen in their video where the dark key are not detected (just imagine what would happen outdoor\sunlight conditions).
also, there was a true statement here that we have developed several generations and technologies at Realsense and I definitely see this as a big advantages real-world usages and HVM requires a lot of experience that only few companies poses and there is a HUGE gap between showing a nice (and environmentally controlled) video.
would be happy to answer further questions,
Roi
What were the comments (in gist?)... Removing any harshness and just scientific.
DeleteHi Roi,
DeleteWhat chipsets will this run on now that Intel seem to be exiting the mobile market with Atom etc? Will Intel be licensing out the platform to run on other ARM units?
low reflective objects are always an encounter to any sort of sensor that uses active illumination. the same is valid for conditions with a high ambient light level. the straight forward answer to both is "use more light". and the "why not?" reasons are ranging from eye safety concerns to power (battery and thermal) downsides.
Deleteits nice being able drive the performance to new limits but using the mentioned nearby options is just not acceptable. the box of options and thus most of the limitations is drawn by the intended application case. asking for breaking the box will render the sensor set invalid for the mentioned case. smart solutions like polarization based 3D scanners might be a big step ahead for some aspects without breaking the box - but even those concept will not be able to solve anything.
maybe you did not understand that a technically impossible item (according to the basic measurement principle: some light gets reflected from the object) is something that will never solve.
for comparing sensor systems it sounds more reasonable to me to look at the performance in the areas where the device should be functional by it's concept.
when it comes to technologies that relies on light coding\modulation than you are right.
Deletedon't forget that stereo is matching a right pixel to the matching left one so its much more robust when its comes to detecting low reflective or high illuminated targets
as for the host platform - we are providing an open-source driver that can be compiled to any host platform (Intel\ARM) and OS (Linux\Win\Android\OSX). as long as it has USB\MIPI
Deletehello Roi, thanks for the response.
Deletestill stereo needs some visible structure on the surfaces. else it can only take edges for depth hints. so its going to be a rough estimator in the critical cases where the scenery is dark. especially curved objects will turn out getting get hard to catch correctly as they dont expose solid outlines but will steadily change shape for different view points.
with a pencil and a piece of paper you should be quickly able to verify whats the difference in detection between a black disc and a black sphere. i would say the sphere will be seen a bit "wider" but on the right depth. the second will not be true for real world curved objects as they can have any sort of curvature not only perfect circles. edgy object instead have much of a chance exposing faces in a very high depth ranges where one camera can see much of it and the other will only see few to nothing. so if you have not enough structure on the surface for detection then detection will suffer and thus the 3-dimensional shape of your scene objects does matter quite much and thus will teach you the (un-avoidable?) precision and detection limits of your device. Anyone operating a 3D sensing device sooner or later came to the conclusion that a quality meter is needed providing areas amongst others with indication of low trust to even no detection. that happy you might be on what is detected - dont forget you will have areas that failed detection. should you worry about the "unknown"? in safety and security applications you definitely should. a system for such a case will have the chance to cease working any now and then whilst using a safe-state strategy that might be sort of a shutdown. the more real world you go the more you will discover your non-detection cases. you will have to put it via the filter of "does it matter to my systems functionality?" - for a device that is compareable to a web cam service level it wont matter in most cases. but you probably dont want to stop there...
the comments are focused to Qualcomm's own performance claims:
ReplyDelete1. number of depth points - they can do 10k vs RS400 which can do 1M
2. depth precision is 0.1mm vs 0.01mm (both are best case I assume)
3. working environment - we are able to work at any ambient environment (Dark->direct sunlight) where SL would have great difficulties. just look at the keyboards of the Qualcomm video and notice that the dark keys are not detectable due to their low reflection.
4. Qualcomm claims the RS is not suited for mobile devices while we are probably the perfect match due to our low form-factor\power and robustness.
hope this answer your question.
Roi
Are you really claiming that depth precision of RS400 is 10 microns?
ReplyDeleteWhat about the accuracy?
this figures are for ideal conditions (for both RS and Qualcomm). in theory, Our highest resolution and wide baseline camera (RS460) can achieve this.
DeleteAccuracy is also a factor of calibration (on top of precision) so it would be definitely higher than the 10micron theoretical number
If practical value is 10 times worse, 100 microns, it would still be a great performarnce. Even 100 times worse, 1mm, is very good.
DeleteI could not find anything related to RS460. Is it different than new coming D415 and D435 cameras?
Depth resolutio of both these solutions won't certainly be 10 or 100 mic's. 10 microns is a ridiculous claim. While I haven't seen the RS400 yet, I'd bet a good amount of money it's not even sub 1 mm in depth resolution. Also the claim about direct sunlight performance is ridiculous too. And not just because the previous Realsense generations were useless with any sunlight cast on the scene.
ReplyDeleteI am willing to make this a public bet for both the sub 1mm and direct sunlight claims you've made.
DeleteI do agree with you about the claims you've made regarding the previous RS generations.
Post some videos, some data, whatever, I am sure everyone would be interested in how you back up a 10 micron claim.
DeleteThis comment has been removed by the author.
ReplyDeleteHey Roi,
ReplyDeleteIf real life precision is 10 times higher than your theoritical claim, this would make 100 microns precision which would still be very precise for such a camera. Or even 100 times worse is still very good performance.
I could not find anything about RS460, is that different than D410 or D435 which are going to be released soon?
have you backed-away from your original bet - "I'd bet a good amount of money it's not even sub 1 mm in depth resolution"?
ReplyDeleteI am willing to post a sub 1mm and to exchange your money if you will forfeit your anonymously :)
You mean you need to have money at risk before proving your claim? I'd say you make the claim, you should provide some data to back it up. Otherwise...
ReplyDelete