a lot of false alarms when applying mixup

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.65k stars 7.96k forks source link

Open Hwijune opened 5 years ago

Hwijune commented 5 years ago

111

hi! @AlexeyAB I recently applied mixup to my yolov3-spp

[net] mixup = 1 blur = 1

After that, there are many false alarms. need more iteration??

aimhabo commented 5 years ago

how many iters did you train after set mixup?

Hwijune commented 5 years ago

how many iters did you train after set mixup?

I trained as much as 20 epoch

Today, the false alarm is reduced a little.

AlexeyAB commented 5 years ago

I trained as much as 20 epoch

What is the size of your dataset? And how many iterations?

Hwijune commented 5 years ago

What is the size of your dataset?

using 700,000 custom data

And how many iterations?

@AlexeyAB more 200,000 iteration batch 64 subdivision 8.

Now the false alarm has been reduced a little, but still a lot.

There are a lot of things that happen when an empty background like first image.

AlexeyAB commented 5 years ago

@hwijune Don't use together these two params

mixup = 1
blur = 1

Hwijune commented 5 years ago

@hwijune Don't use together these two params
mixup = 1
blur = 1

thank you, I'll try.

tuongtranngoc commented 4 years ago

@AlexeyAB and @hwwwi1 what do that mixup do in data augment?

stephanecharette commented 4 years ago

@AlexeyAB and @hwwwi1 what do that mixup do in data augment?