3 ms·

The sad thing is that Intel had a real advantage in RealSense... they had good processors in there (especially with the Movidius chips) that were well integrate

The sad thing is that Intel had a real advantage in RealSense... they had good processors in there (especially with the Movidius chips) that were well integrated (e.g. they didn't have to support a fancy SDK for all users - just the internal Realsense group). Doing the same 3d processing with ARM or Intel is 1/10th the frame rate at much more power.

reasonabl_human 5y ago
Can you elaborate on the benefits of what intel built here as opposed to doing the processing on host machines? My company uses these in some products but I never investigated the market or trade space for this type of device
- morcheeba 5y ago
  Yeah! There is a lot of processing to do the disparity calculation... it's basically doing lots of correlations with a different amount of x shift and finding the delta-x that has the highest correlation. This delta-x disparity gets calculated in to a z distance. So, intel made a custom processor for this that's really good at the correlations. The original one was the D4, and later they used their movidius chip. Both have lots of multiplier-accumulate silicon, so it can do the computations in parallel. Their architectures are also set up for convolution (which re-uses a lot of data) rather than random-processing (like a CPU does), so they could feed these math engines without a lot of data transfer -- this makes it more power efficient. You could do something similar in an FPGA, but dedicated silicon is going to be faster, cheaper, and use less power.
snovv_crash 5y ago
Thing is, my personal perspective (haha) is that single-shot stereo is a dead-end tech. There simply isn't enough information to infer depth in the difficult situations, range is severely limited, and every frame it basically re-calculates all of the information from the previous frame from scratch, making it wildly inefficient. Of course, it works well as a tech demo and in controlled environments, and it's easy to understand and then brute-force with custom hardware. The problem is that it sucked the life out of better techniques which might actually work in the real world.

reasonabl_human 5y ago

Can you elaborate on the benefits of what intel built here as opposed to doing the processing on host machines? My company uses these in some products but I never investigated the market or trade space for this type of device

morcheeba 5y ago
Yeah! There is a lot of processing to do the disparity calculation... it's basically doing lots of correlations with a different amount of x shift and finding the delta-x that has the highest correlation. This delta-x disparity gets calculated in to a z distance. So, intel made a custom processor for this that's really good at the correlations. The original one was the D4, and later they used their movidius chip. Both have lots of multiplier-accumulate silicon, so it can do the computations in parallel. Their architectures are also set up for convolution (which re-uses a lot of data) rather than random-processing (like a CPU does), so they could feed these math engines without a lot of data transfer -- this makes it more power efficient. You could do something similar in an FPGA, but dedicated silicon is going to be faster, cheaper, and use less power.

snovv_crash 5y ago

Thing is, my personal perspective (haha) is that single-shot stereo is a dead-end tech. There simply isn't enough information to infer depth in the difficult situations, range is severely limited, and every frame it basically re-calculates all of the information from the previous frame from scratch, making it wildly inefficient. Of course, it works well as a tech demo and in controlled environments, and it's easy to understand and then brute-force with custom hardware. The problem is that it sucked the life out of better techniques which might actually work in the real world.