search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.68k stars 7.96k forks source link

Questions about YOLOv4 #5346

Open stephanecharette opened 4 years ago

stephanecharette commented 4 years ago

I see there is now a YOLOv4! This is exciting. Can someone comment on some of the highlights/differences between v3 and v4? The readme doesn't have much details.

When setting up new projects, should I default to YOLOv4 instead of v3? Worded differently: are there reasons why I would use v3 instead of v4?

Most of the darknet projects I've done in the past have been YOLOv3-tiny. But it looks like there is no such thing when it comes to v4. Is this in the works?

Lastly, I see 2 v4 .cfg files. The only real difference between them is learning_rate=0.001 and learning_rate=0.00261. How do we choose which one to use?

AlexeyAB commented 4 years ago

When setting up new projects, should I default to YOLOv4 instead of v3?

Yes.

Worded differently: are there reasons why I would use v3 instead of v4?

Only if you want to use other frameworks (TF, Pytorch, ...) which don't support YOLOv4 yet.

Lastly, I see 2 v4 .cfg files. The only real difference between them is learning_rate=0.001 and learning_rate=0.00261. How do we choose which one to use?

Since yolov4.cfg may be unstable for some other datasets, we provide yolov4-custom.cfg with slightly lower learninig_rate. I think we will fix this issue by using max_delta=5 in the [yolo] layers instead of decreasingf learninig_rate.

stephanecharette commented 4 years ago

Thanks, Alexey. I see that YOLOv4 has 162 layers. Compared to 107 in YOLOv3 and 24 layers in YOLOv3-Tiny. Especially when going from v3-Tiny, that is quite the jump. I haven't tried to retrain any of my existing networks in v4 yet, but I suspect that processing time per frame will increase significantly when going from v3-Tiny to v4. Several of my neural networks run on small devices, like Jetson Nano, which have limited computing power.

Are there plans on defining a v4-Tiny configuration?

AlexeyAB commented 4 years ago

Are there plans on defining a v4-Tiny configuration?

Yes.

znsoftm commented 4 years ago

Looking forward to your v4-tiny configuration.

znsoftm commented 4 years ago

So when is the exciting time?

sealedtx commented 4 years ago

+1! Waiting for v4-tiny, thank you @AlexeyAB

AlexeyAB commented 4 years ago

YOLOv4-tiny released: 40.2% AP50, 371 FPS (GTX 1080 Ti)

cmp