search

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
341 stars 97 forks source link

Possible to run on a Smartphone? #93

Closed d-lay closed 3 years ago

d-lay commented 3 years ago

Hi, I was wondering if it might be possible to run your 6D Pose Estimator even on the limited hardware of a smartphone. Before diving too deep into it, I wanted to get the opinion of more experienced people in the field.

If you think it is possible, is a rough estimation for fps possible (Lets assume it would run on an Adreno 650 mobile GPU)?

If you think it is not possible could you briefly elaborate on the reasons.

Any insights are appreciated. Thank you very much!

MartinSmeyer commented 3 years ago

I have not tried it on a smartphone, but I tried it on a Jetson TX2 where you can run it in real-time. The model size can be reduced by changing the NUM_FILTER parameter in the training config. The codebook requires around 40MB in fp32 but could also be reduced.

d-lay commented 3 years ago

Thank you for your reply! That is already very valuable information and it seems more promising than I anticipated.

The AI performance of the Adreno 650 mobile GPU (1.2 TFLOPS) is comparable to the Jetson TX2 4GB Module (1.3 TFLOPS). And 40MB RAM is doable on an Android.
Is this already enough to assume, that it is possible in (near) real-time on a smartphone? Or do you think other hardware limitations are critical? Like having significantly fewer cores on the smartphone and limits of the CPU.

MartinSmeyer commented 3 years ago

No I think your assumptions are valid. Keep in mind that this is per object though. I ran it on a couple of objects in parallel on the tx2