search

youmi-zym / GO-SLAM

[ICCV2023] GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction
Apache License 2.0
344 stars 30 forks source link

How to train on custom datasets #35

Open michaelw-0615 opened 3 months ago

michaelw-0615 commented 3 months ago

Hi authors,

Thanks for your excellent work.

I've noticed that this repository includes the checkpoint of a pretrained model ("./pretrained/droid.pth"), could you provide some details about which datasets this checkpoint file is based on and how this model is trained?

Also, if I plan to train on custom monocular datasets like Tanks and Temples to obtain a new set of checkpoints, how could I access the training scripts and what kinds of data (RGB maps, depth maps, poses, etc.) are required for training? Thank you!

Best, Michael W

youmi-zym commented 2 months ago

Hi,

The frontend is pretrained and fixed when doing SLAM things. So, if you want to deploy to another dataset, it doesn't need to train the frontend network again.

But, if you want to improve/finetune our pretrained frontend network, you can follow the steps mentioned in DroidSLAM.

Best, Youmin