Darknet: I'm going to transplant the yolov3-mini to a Rasberry pi3B+.Search for some advices.

[ZyuZyuZyu]

I'm going to transplant the yolov3-mini to a Rasberry pi3B+ with one or two intel movidius neural sticks for real-time object detection. And I'm wondering how high the FPS I can reach the most. Has anyone done this before? I am looking for some advices, or literatures, or documents. Thanks a lot!

[jinyu121]

FYI:

| Version | Network | Speed | Comment |
| --- | --- | --- | --- |
| pjreddie | yolo v2 | 300+s | Old version|
| pjreddie | tiny yolo v2 | - | Weights error on Pi |
| AlexeyAB | yolo v3 | - | Forget... Sorry |
| AlxeyAB | tiny yolo v3 | 15s | |
| AlxeyAB | tiny yolo v3 | 11s | OPENMP=1 |
| AlxeyAB | tiny yolo v3 | 7.3s | OPENMP=1 && -marm|
| darknet-nnpack | tiny yolo v2 | 1s | Weights error on Pi |
| darknet-nnpack | tiny yolo v2 | 0.8s | OPENMP=1 && -marm|
| darkflow | tiny yolo v2 | 100++s | Forget... Sorry |
| darkflow | yolo v2 | 30min | |

Origin: https://haoyu.love/blog511.html

You can try -mfpu=neon-vfpv4 -funsafe-math-optimizations -ftree-vectorize on this repo just like https://github.com/digitalbrain79/darknet-nnpack/blob/master/Makefile

More info about weights error: the seen is 64 bit in weights file, but Pi is 32-bit. So the weights are wrong. This phenomenon only exists in tiny-yolo-v2. I forget which issue we have discussed.

More info about pjreddie's version: The latest version can not work in both detection and classification, since there is an error just before output the result. https://github.com/pjreddie/darknet/issues/823

---

Oh, these data are just pure darknet, without intel neural sticks. Waiting for your data~

[shizukachan]

Updated darknet-nnpack on yolov3-tiny-voc runs 1.1s for first inference, 0.73s for subsequent inference.

See https://github.com/shizukachan/darknet-nnpack for code.

[ZyuZyuZyu]

@jinyu121 @shizukachan Thanks~

[caihaocong]

I'm going to try transplant the yolov3 to intel movidius neural stick for detecting classes too. We can community with each other. My email is [email protected].

[ZyuZyuZyu]

@caihaocong I have mailed you.Thanks!

[spinoza1791]

Updated darknet-nnpack on yolov3-tiny-voc runs 1.1s for first inference, 0.73s for subsequent inference.

See https://github.com/shizukachan/darknet-nnpack for code.

Please add xnor to your repo for 4x speed-up!