Custom dataset consisting only of syntheised images

[AminSeffo]

Hey everyone,

I would like to know if anyone here trains his custom dataset only with synthesized images without any real images.
I did it already and used the same config, which is already used in Detectorn2 tutorial but the result is not sufficient. Recently I found a paper https://www.arxiv-vanity.com/papers/1710.10710/ that presents a new method to boost performance during training by freezing the feature extractor in all layers.
So do someone know, how can I do this in detectron2 or has other methods to improve the performance?

Any answer is very appreciated.

[zhanghang1989]

You may freeze different number of stages for the backbone. If you'd like to freeze the entire feature extractor, set it to 5 https://github.com/facebookresearch/detectron2/blob/main/detectron2/config/defaults.py#L137

[AminSeffo]

@zhanghang1989 thank you very much for your response. Sorry because the above link doesn't work, here is the correct one: https://arxiv.org/abs/1710.10710
Would this also mean that the feature extractor’s weights are frozen and they can't adapt during the training?

[zhanghang1989]

Yes, frozen weights won't be adapted