Darknet: Effect of Training Images Missing Labels

[nseidl]

Hi.

I have some images that contain multiple sofas.
Imagine an image that has 3 sofas: sofa1, sofa2, sofa3.

The .txt label file corresponding to the .jpg only contains one line: the box information for sofa1.

Will this negatively effect the training process and the model?

[kossolax]

yes

[nseidl]

Would anyone mind elaborating?

[kossolax]

if any annotation is missing on the image, the training process will think it's a false positive and thus, will try to not predict it.

[Leo16112]

Hm.. for example, I have 2 classes, 200 labels for each. And in one image I changed the labels. Object one got label of 2 and object two label of 1. Can I consider that my detection will loose 0.5% of efficiency or is it wrong calculation?

[kossolax]

That's an hard question. but I guess there is possible 3 scenarios. You might see these scenario by playing with this: https://playground.tensorflow.org but keep in mind that problem still very basic. Object detection is not. The learning process has some strategies to avoid noise, but the more noise you add, the more difficult it is. I've seen some model converging with high noise, but the conf given was between 0.1-0.01%.

• few missing labels, few mAP reduction because validation also has missing labels. Model will still predict it. - https://playground.tensorflow.org/#dataset=spiral&noise=1
• some missing labels, much mAP reduction, because model will try not to predict them. https://playground.tensorflow.org/#dataset=spiral&noise=10
• much missing labels, broken mAP, because model cannot understand what the class is. https://playground.tensorflow.org/#dataset=spiral&noise=100