Evaluate upcoming Jetson Xavier for OpenDataCam about opendatacam

[b-g]

@florianporada Thanks for the assessment! Agreed! :) We should buy one once it becomes available. I signed up for a notification at nvidia ... let's wait.

[tdurand]

First benchmark of Jetson Xavier 👀

Running YOLO

Our YOLO implem "mesos" fork doesn't work out of the box on Xavier, running it leads to a Seg fault 😐 , might be because of opencv, might be because of CUDA or any other dependency, not idea on why is failing, error doesn't give any info

After a bit of research, found out that the current "state of the art" implem of YOLO (v3) seems to be this repo: https://github.com/AlexeyAB/darknet , which has improvement over the "official" pjreddie/darknet one regarding specific architecture like Jetson Xavier or other GPUs hardware

So I went ahead, and after compiling and running it, turns out performances are really good 🎉, significantly faster than the TX2, seeing a 4-10x perf boost :

YOLOv2 & VOC (same as current open data cam)

Power mode	FPS
Default (15W)	25 FPS
Max out (30W+)	30+ FPS (webcam steam can't go higher)

We were getting max 7 FPS with TX2

YOLOv3 & COCO

Power mode	FPS
Default (15W)	11.5 FPS
Max out (30W+)	22.5 FPS

We were getting max 2 FPS with TX2 on YOLOv2 + COCO

FFServer + YOLO v3

From a really quick benchmark, seems to be possible without much perf hit ( 1 FPS penalty..), see full notes for more on this

Full notes

I wrote more notes on all the setup + perf exploration: https://docs.google.com/document/d/1oDk7Nf95Pm2eJqHsC2TL24v87CsHW8wtENbJlkBSmNE/edit#heading=h.6px293e6d645

Further potential exploration:

• Try to run this new darknet implementation on Jetson TX2 to see if perfs improves
• Test a lighter implementation of yolov3: https://github.com/AlexeyAB/yolo2_light , claim +30% speedup for only -1% "accuracy"

Up next

I think this is promising as perfs are significantly better this can enable:

• a live preview of counting / detection to improve UX
• potential better detection and tracking, more FPS + Yolov3 trained with COCO dataset should be better at detecting things and more reliable

But we won't avoid to update our darknet implementation / port the websocket feature if we want to go further with Jetson Xavier... I'm not sure of the complexity of this but I remember looking in the C code done on the mesos fork and it didn't seem much.. but things are always more complex than they seem, and I don't have much expertise on both darknet source code and C code..

[florianporada]

If 20x the performance of its predecessor statement is true we could use higher resolution video footage and a higher framerate to boost accuracy. (maybe use a better trained model as well).
Also the base-system will be Ubuntu+JetPack which should provide us with the necessary libraries to run the code. (if the versions are compatible).

One downside is, that the board does not come with build-in WiFi anymore. So the parts (usb-dongle) we need to build the project will increase (+ another spof).

I'm in for testing it and doing a port if we get one of the boards in our hands 😜

[b-g]

Order! Shipping Sep 20 2018, should arrive beginning of Oct.