Destination of SDF map and resulting mesh when using ROS implementation.

[gsznaier]

Hi!

Thank you for sharing your code, this is very cool work!

I would like to compare against your approach and I had a quick question regarding your implementation when running with ROS. Where does your approach store the resulting SDF map and mesh as it trains?

Thanks in advance!!

[joeaortiz]

Hi Gadi,

The SDF map is simply stored in the weights of the MLP network which is in the iSDF ROS node. The mesh is generated by querying the SDF network on a regular grid and then running marching cubes (https://github.com/facebookresearch/iSDF/blob/main/isdf/modules/trainer.py#L1500). The mesh is only generated when the visualisation window asks for a mesh to display.

Feel free to reopen if you have follow up questions.

[gsznaier]

Hi Joe,

Thanks for the quick response! One more question, does the SDF store the network weights in a specific folder during training or is the result lost once the training script is terminated?

Thanks in advance!

[joeaortiz]

You can save the weights by saving the model state dict, e.g. torch.save(model.state_dict(), PATH). You can set the period at which the weights are saved in the config file: https://github.com/facebookresearch/iSDF/blob/main/isdf/train/configs/replicaCAD.json#L27