Question of results when learning TAO with incorrect labeling data

[user136657]

Hello, I am posting a question during TAO learning.

I wonder if the learning works normally when I do TAO learning when there are more than two labeled objects in the labeling data or when the range of boxes in the labeling is outside the range of the image.

I understand that when learning ultralytics yolo, when there is such a problem, the learning is conducted by excluding the corresponding labeling.

Is there such a thing in TAO?

[Morganh]

@user136657

there are more than two labeled objects in the labeling data or when the range of boxes in the labeling is outside the range of the image.

Could you please share one or two examples?

[user136657]

• first_labels.txt ( more than two labeled objects in the labeling data )

person 0.00 0 -10.00 300 200 400 500 -1.00 -1.00 -1.00 -1000.00 -1000.00 -1000.00 -10.00
person 0.00 0 -10.00 300 200 400 500 -1.00 -1.00 -1.00 -1000.00 -1000.00 -1000.00 -10.00

• second_labels.txt (range of boxes in the labeling is outside the range of the image - image size 1920 1080)

person 0.00 0 -10.00 300 200 3000 5000 -1.00 -1.00 -1.00 -1000.00 -1000.00 -1000.00 -10.00

I can give an example like the above

[Morganh]

In YOLOv4, there is a parameter as below, which can ignore this difficult boxes.

• include_difficult_in_training: A flag specifying whether to include difficult boxes in training. If set to false, difficult boxes will be ignored. Difficult boxes are those with non-zero occlusion levels in KITTI labels.

[user136657]

@Morganh

include_difficult_in_training: A flag specifying whether to include difficult boxes in training. If set to false, difficult boxes will be ignored. Difficult boxes are those with non-zero occlusion levels in KITTI labels.

Is it applicable to detectnet_v2 other than yolov4?

[Morganh]

For detectnet_v2, there is not.
All YOLO series networks have include_difficult_in_training.

Suggest you to use YOLOv4_tiny instead.