Limit bounding boxes

[amandakj]

Hi.

I have images with three different objects, where they only appear once. When I am training and detecting I sometimes get multiple bounding boxes for each class. I want to limit the number of bounding boxes to one per class.

Is there a hyperparameter or a function I can change to easily fix this problem?

[glenn-jocher]

@amandakj yes, you can control this by varying NMS settings to suit your needs:

• Agnostic NMS will run NMS across all classes at once, eliminating incidences of overlapping boxes from different classes (False by default). To enable: python detect.py --agnostic
• Multilabel NMS assigns all classes above threshold to the same bounding box. Enabled by default. You can disable here:

  yolov5/utils/general.py

  Line 409 in 9d87307

  multi_label = nc > 1 # multiple labels per box (adds 0.5ms/img)

[amandakj]

Thanks, @glenn-jocher! 😄

[alucor-it]

Hi @glenn-jocher ,
I have set --agnostic-nms but I still get multiple bounding boxes per class in my saved text file. Do I have to make any other changes?

[glenn-jocher]

@alucor-it multiple detections within one class in an image is not incompatible with agnostic NMS. See https://www.quora.com/What-does-%E2%80%9Cclass-agnostic%E2%80%9D-in-most-of-the-object-detection-papers-mean

[glenn-jocher]

@amandakj we are modifying the default inference behavior here to multi_label=False here after multiple user feedback. See #2252 for details.

[jeannotes]

1. I think the multilabel only reference to those different objects but close to each other and overlap closely, right?
2. but when enable multi-label, what is the insight, meaningful?

[glenn-jocher]

@jeannotes this is for use cases that require this configuration.

[jeannotes]

@glenn-jocher why needs this requirements?

[glenn-jocher]

@jeannotes it's not a requirement.

[SpongeBab]

@jeannotes ,it not must, it just need or not.

[jeannotes]

@glenn-jocher @SpongeBab but I still don't understand that when set "multi_label = nc > 1 ", what is the advantages

[myasser63]

@glenn-jocher How can I set yolov5 for multiclass not multilabel task? is this done by Agnostic NMS only or there's another setting?

[myasser63]

@glenn-jocher Do I need to change Cls Loss from BCEwithlogits to CrossEntropy loss to convert to MultiClass ?

[glenn-jocher]

@myasser63 multi-label NMS will allow for multiple classes on a single box:

yolov5/utils/general.py

Lines 806 to 813 in d7bc5d7

	def non_max_suppression(prediction,
	conf_thres=0.25,
	iou_thres=0.45,
	classes=None,
	agnostic=False,
	multi_label=False,
	labels=(),
	max_det=300):

[myasser63]

@glenn-jocher Sorry I don't fully understand. My dataset is having 4 classes mutually execlusive. From my understanding BCEwithlogits is for Mulitlabel problems. So to make only the task for multiclass only, just make multi_label to False without changing the loss function?. I mean during training and validation not inference.

[glenn-jocher]

@myasser63 yes BCE is multi-label capable, but by default multi-label is False and NMS is set to output only 1 label per box.