search

VisDrone / DroneVehicle

Drone-based RGB-Infrared Cross-Modality Vehicle Detection via Uncertainty-Aware Learning
417 stars 46 forks source link

Inconsistent labeling #48

Open love-whut opened 1 month ago

love-whut commented 1 month ago

Hello, I found that many images in the test set have inconsistent RGB and infrared labels. Which one should I follow?

OBrink commented 1 week ago

What exactly do you mean here? What exactly is inconsistent?

I have found that the class names in the provided annotations are inconsistent. For example, "freight_car" is referred to as "feright_car", "feright" or "feright car".

love-whut commented 1 week ago

What I mean is that the labels of RGB images and IR images in the test set are inconsistent.

OBrink commented 1 week ago

seems related to #47. Do you mean that objects are not annotated in the very dark RGB images?

love-whut commented 1 week ago

yes

OBrink commented 1 week ago

That seems to make sense to me though. I would not expect that regions are annotated when you cannot recognise them in the dark images. I think this would lead to problems for detection models that try to learn based on the RGB data. It's even explicitly presented like that in the README of this repo (right image):

image

Consequently, if you want to train a multimodal detection model, you would probably want to follow the IR data as they contain the complete annotations. In that case, the RGB data does not contain the information necessary to find the objects.