custom dataset size for segmenatation


I was following mobilenetV2 notebook from the tutorial, and I added my custom dataset(limes).
I found that with around 300 images it only has 65% accuracy, I wonder if that is because of the size of the dataset, or if there was any problem with merging my dataset into COCO.
If mobilenetv2 is not the best model for this, is there any recommendation for segmentation? I tried Detecton2 but it does not support onnx runtime.


  • سعد الفريدي
    سعد الفريدي Posts: 447 🔥🔥🔥

    Hey there,
    Try to reduce the number of images,

    Because the number may be the reason for that

  • GuuInTheHaus

    Hello سعد الفريدي,
    the pizza example has ~1800 images for training. I wonder if ~300 is already too big?