Training with synthetic data

[DeepBernhard]

Dear Deep Learning Pros!

I am trying to refine the sample object-detection network with synthetic data instead of doing it with the KITTI data as in the tutorial. My goal is to compare it to a model refined with the KITTI data on real data. In the first try I used all of my generated synthetic data (19000 images and labels) for training and the original KITTI data for validation. It is still in progress, but it looks like it's going to suck. Should I use synthetic data for validation instead and only compare the performance on real data? What do you think is the best way to refine networks with synthetic data?

Sorry for asking silly questions. Thx Bernhard

[DeepBernhard]

In the end I used a dataset with synthetic data for training and validation which improved the training process:

After the successful training I tested the resulting model against 100 sample KITTI validation pictures and got 158 moslty right positioned bounding boxes but also some false positives. The KITTI data refined model found on the same pictures 294 bounding boxes, some of which were duplicates on the same object.

Sample result from the synthetic refined model:

Sample result from the KITTI refined model:

We were satisfied with the result of our first try, and there is a lot of room for improvement.