Training on simulation and testing on real-world benchmark

[nikola310]

Thank you for sharing your interesting work. Have you perhaps tried training your model on GTA data set and testing on real-world images? It would be interesting to see how model trained on synthetic data responds to real-world data.

[eborboihuc]

Thanks for your interest in our work!
We focus more on combining quasi-dense similarity learning and 3D instance dynamics, to track moving objects in a 3D world. But, we also notice the potential of using our GTA dataset as a data source to transfer the ability on real-world data. It's a great topic to discuss with others if anyone happens to have some experiments about it.