yolov9 icon indicating copy to clipboard operation
yolov9 copied to clipboard

Published 20 hours ago • verified publisher icon WongKinYiu

why is YOLO-e performing worse than YOLO-c on my custom dataset?

Open lucysumi opened this issue 10 months ago • 6 comments

I run YOLOv9-c and YOLOv9-e. YOLOv9-c results for prec, rec, and map values are 98.9, 98, 99.2 and 48.7 respectively while for YOLOv9-e results were 95.8, 95, 94.4, 45.7.

wasn't E supposed to give better results than C? Thats what I assumed or what am i missing here ? @WongKinYiu Please help me understand if I'm missing out anything?

lucysumi avatar Apr 10 '24 06:04 lucysumi

It depends on your dataset, the task to be accomplished, the number of images you have, and their diversity.

When you have little information in a piece of data, a large model will abstract it too much, and therefore have bad results, for example.

Youho99 avatar Apr 10 '24 09:04 Youho99

Thanks for reply. My dataset is a sequence of CCTV footage (short clips) with a man carrying an object. There is no variety in it. I just extracted frames (images) from those clips and used these images for training. Just 500 images. That could be the reason.

lucysumi avatar Apr 10 '24 10:04 lucysumi

If I have a dataset contains about 2,000 images, could I see it as a large dataset?

TJ20GJQ avatar Apr 24 '24 07:04 TJ20GJQ

If I have a dataset contains about 2,000 images, could I see it as a large dataset?

I think no But it depends How many classes / instances per classes do you have ?

Youho99 avatar Apr 24 '24 08:04 Youho99

If I have a dataset contains about 2,000 images, could I see it as a large dataset?

I think no But it depends How many classes / instances per classes do you have ?

I have 10 classes and about 40,000 instances totally. I think it is large, so I plan to use yolov9-e-dseg.yaml to solve it. Maybe I should try both and compare them.

TJ20GJQ avatar Apr 24 '24 08:04 TJ20GJQ

If I have a dataset contains about 2,000 images, could I see it as a large dataset?

I think no But it depends How many classes / instances per classes do you have ?

I have 10 classes and about 40,000 instances totally. I think it is large, so I plan to use yolov9-e-dseg.yaml to solve it. Maybe I should try both and compare them.

Yes, do some experiment plan

Check this for more informations (yolov5 docs, but it's pretty the samedi things) : https://docs.ultralytics.com/yolov5/tutorials/tips_for_best_training_results/

Youho99 avatar Apr 24 '24 08:04 Youho99