Can loftr be trained directly on the data in the dataset?

[0121910870533]

Can you give us more details

[RingoWRW]

We provide localization dataset, but you can use render database(rgb/depth/pose/intrinsic) to train loftr.
First, you should compute their overlap and generate .npy file. Second, write dataloader.py as megadepth.
We provide a localization dataset for your use. To train the LoFTR model, please follow these steps:
##Data Preprocessing:

1. Compute image overlaps using the render database (including RGB images, depth maps, camera poses, and intrinsic parameters)
2. Generate corresponding .npy files containing the preprocessed data

##Dataloader Implementation:
Implement a dataloader.py script following the same structure as the MegaDepth dataset loader