0
$\begingroup$

Suppose we want to train a model to detect various objects. Let's say we have training data of those objects in various backgrounds along with their bounding boxes. Basically these objects have been three dimensionally created and the bounding boxes have been drawn on them. Then these have been "synthetically inserted" into various blank backgrounds.

Why would a model trained only on this data do better than a model that has this data along with "real" data of these objects with their bounding boxes manually drawn?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

I am not absolutely sure, but I guess it is due to the domain gap. As far as I have seen in my project where YOLOv3 was trained on synthetic images, it performed better when the model was trained and tested on the same domain (synthetic), while the performance drops when we introduce real images for testing. So when you include real images along with synthetic images, you might have to use some domain adaptation methods to improve the performance.

Improve this answer
$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .