Optimal dataset

hi
i made a dataset it was better than before and mAP increased from 19% to 59% but it still low to me, the precision is 85% almost but the recall is low 56%
I need to make mAP almost 97%
011
012
013
019
010
please i need the optimal dataset also i need to understand how to make backgrounds, with dataset, i made 1000 image with no Annotate i mean didn’t put any labels in it is that Backgrounds?

  • Images per class. ≥ 1500 images per class recommended
  • Instances per class. ≥ 10000 instances (labeled objects) per class recommended
    1500 per class i dont have that amount of images for each class,
    what does this mean 10000 instances for each class?

Hi @mo_t1 , and great job!

Is the project requirement to achieve 97% or more on mAP? As that will take time, and will be easier with fewer classes. Gun detection is just a very difficult problem as guns look very different based on the manufacturer, and weapon type (rifle, pistol, etc.)

Yes, you are correct, images in the dataset without any labeled items are referred to as “Null” or “Background” images: The Difference Between Missing and Null Annotations

Here is how to create a synthetic dataset (places annotated image examples on random background):

You can also search Roboflow Universe for more datasets: https://universe.roboflow.com

And continue with Active Learning to include examples of guns that are not in your dataset splits (train/valid/test), and reserved for testing.

You can download more images to test, or add to the dataset, from Roboflow Universe after searching “guns”: https://universe.roboflow.com

1 Like