Datasets with "Background Images"

[AlexanderJGomez]

Hello everyone,

I've finetuned a model on some custom data and am getting many false positives and had been pondering whether there might be a way to introduce "negative" examples that basically provide a way to tell the model that the image does not contain our objects. So, for example, if I trained a one class detection model for goldfish that was getting false positives in images containing beta fish, I might include images with just beta fish in my dataset in order to help reduce the false positives from the similar looking betafish.

I have been unable to find the proper terminology for what this is called but I did see it referred to as a "background image" in this post about improving model performance.

Does anyone know where I can find more details on this concept? and how can I introduce this kind of data into training with YOLOv5?

Thanks!

[glenn-jocher]

@AlexanderJGomez yes that's correct, introducing background images will help reduce False Positives (FPs). You can introduce FPs into your dataset simply by putting them in your train: and val: directories (example from voc.yaml shown below). No labels *.txt files are required and no other action is required, YOLOv5 will detect and train with the additional background images next time you train.

yolov5/data/voc.yaml

Lines 12 to 15 in 2a28ef3

	# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
	train: ../VOC/images/train/ # 16551 images
	val: ../VOC/images/val/ # 4952 images

[AlexanderJGomez]

Perfect! @glenn-jocher thanks for the quick response.

Side note: very much appreciate all the work you're doing with the repository. I hope you continue the great work