Closed CedricVandelaer closed 4 years ago
The problem is the top view.. the 3D network is not trained on such views for it work.. I would just use the 2D detections and then fit a model using EPNP. Another way to address this problem is to retrain the network on this data with top views. We are currently working on improving the 3D reconstruction in future versions of the code. Will release it once the work gets published.
Hello,
First of all thanks for creating and sharing this work. I have a question though. I have used the code on some traffic intersection images we have collected in Belgium. The output is pretty good but we only got 2D outputs.
I have then enabled the KGNN3D setting in the config file, but now the outputs get really bad (see before and after picture). I was wondering if you have any idea why this is the case or how to improve this? I used the pretrained weights, but maybe the network wasn't yet trained using the 3D head?
Without 3D:
With 3D:
Best regards, Cédric