[D] YOLOv4 is too biased

[saikrishnadas]

Goal : To train a model that can detect 70 different Classes of fruits and vegetables that look similar.

I trained these models with YOLOv4 and the model was biased to 2-3 classes. The object detected wrongly doesn't look similar at all. ( eg. The model detected banana as apple ) .And sometimes no detection at all. I cross-checked the dataset and annotations, everything was alright.

Can anyone help me why the model is too biased? @AlexeyAB

[AlexeyAB]

@saikrishnadas

• Show 2-3 training examples with bounded boxes

The object detected wrongly doesn't look similar at all.

• Show 2-3 detection examples with bad bounded boxes

• Attach your cfg-file

[saikrishnadas]

Few training images ( Capsicum Yellow and Capsicum Green):

Image that was tested:

tested on postman, result: { "CameraID": "1", "id": [ "Capsicum_Yellow10000209", "Capsicum_Yellow10000209", "Capsicum_Yellow10000209" ], "score": [ "99.81", "99.80", "99.72" ], "count": 3 }

{ "CameraID": "1", "id": [ "Capsicum_Yellow10000209", "Capsicum_Yellow10000209", "Capsicum_Yellow10000209", "Capsicum_Yellow10000209" ], "score": [ "99.93", "99.93", "99.88", "99.63" ], "count": 4 }

This is an example, the wrong prediction occurs to other objects too.

Total number of classes trained: 70

Cfg file used:

custom-yolov4-detector.zip

[AlexeyAB]

Use for training in cfg-file (disable color data augmentation):

 saturation = 1.1 
 exposure = 1.5 
 hue=0.0

instead of https://github.com/AlexeyAB/darknet/blob/560fda34947fe842805b49a1eeda8f2c162cdca9/cfg/yolov4x-mish.cfg#L14-L16

More about it: https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-%5Bnet%5D-section

[saikrishnadas]

Thanks for the response. Let me try it that way. And will get back to you with the result

[Goru1890]

I'm planning a training, yolov4 with two classes: male and female. The only way two be able to differentiate a male from a female is through the colour of the feathers. Should I use the same parameters as above?

[AlexeyAB]

Oh, there were mistake, use

 saturation = 1.1 
 exposure = 1.5 
 hue=0.0

instead of https://github.com/AlexeyAB/darknet/blob/560fda34947fe842805b49a1eeda8f2c162cdca9/cfg/yolov4x-mish.cfg#L14-L16

Should I use the same parameters as above?

Yes.

[saikrishnadas]

Damn! I started my 2,50,000 with expo 1.1 and Hue .0 as you said before.

On Thu, 3 Dec 2020, 7:45 pm Alexey, notifications@github.com wrote:

Oh, there were mistake, use

saturation = 1.1 exposure = 1.5 hue=0.0

https://github.com/AlexeyAB/darknet/blob/560fda34947fe842805b49a1eeda8f2c162cdca9/cfg/yolov4x-mish.cfg#L14-L16

Should I use the same parameters as above?

Yes.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/AlexeyAB/darknet/issues/7065#issuecomment-738020239, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHLIRA4DI5JAF63DDVM6HC3SS6MP3ANCNFSM4UKHOYJA .

[stephanecharette]

If you're looking to get a better understand of those values, @saikrishnadas, see this page: https://www.ccoderun.ca/darkmark/DataAugmentationColour.html

[saikrishnadas]

I get a very low map rate. I usually get 89-90% at this epoch range.

(next mAP calculation at 135322 iterations) Last accuracy mAP@0.5 = 8.62 %, best = 8.90 % 135248: 1.293633, 1.071062 avg loss, 0.002608 rate, 1.879425 seconds, 51935232images, 115.112189 hours left Loaded: 0.000055 seconds

[saikrishnadas]

And a few quick questions,

• Making the saturation = 0.0 , exposure = 0.0 , hue=.0 will train my image as it is? Will that improve my detection in my case?
• I use 8 GPUs for training, so I changed my learning rate as suggested in Github. So, does that make my detection better, or reducing the learning rate (value/8) will make my model learn less.?

@stephanecharette

[stephanecharette]

• Making the saturation = 0.0 , exposure = 0.0 , hue=.0 will train my image as it is?

Yes.

• Will that improve my detection in my case?

That I don't know. Trial and error, you'll have to try different things and see. It really depends on the images in your training set, and the actual images you are using for inference.

[saikrishnadas]

I get a very low map rate. I usually get 89-90% at this epoch range.

(next mAP calculation at 135322 iterations) Last accuracy mAP@0.5 = 8.62 %, best = 8.90 % 135248: 1.293633, 1.071062 avg loss, 0.002608 rate, 1.879425 seconds, 51935232images, 115.112189 hours left Loaded: 0.000055 seconds

@stephanecharette

[saikrishnadas]

I need support with the above problem. @AlexeyAB

[AlexeyAB]

• Attach your cfg-file and chart.png file.
• What command do you use for training?