Problem with low metrics

[panagiotamoraiti]

Hello,

I've been working on running YOLOv9 on my custom dataset with 10 classes. After repeatedly training with poor results, metrics consistently below 1%, I discovered that reducing the learning rate to 0.0001 and disabling most data augmentations significantly improved performance. The only augmentation I kept was a horizontal flip with a probability of 0.2.

Since making these changes, my metrics have started to improve. After just 10 epochs, I'm now seeing around 50% AP and AR. If you're struggling with low metrics, I highly recommend trying these adjustments.

[Rhroan]

My experence ， when few class , use Mosaic and RandomCrop is deficult to train especially when there is only one category.

I suggest adding
RandomBrightness,
RandomContrast,
RandomSaturation to image enhancement training