Usage with custom dataset

[jrdimale]

Thanks for the amazing work!

I have a set of monocular images captured from a drone’s point of view (covering large outdoor scenes), and I’d like to train and evaluate a model using this custom dataset.

Could you share any insights on how to best format the data and what preprocessing steps are needed in order to replicate the Aachen experiment with my own data?

[xjiangan]

Hi! Thanks for your interest in our work — we're excited to hear you're trying it out with your own drone-captured dataset!

To experiment with custom data, you'll need your monocular images along with their corresponding camera intrinsics and extrinsics. These typically can be obtained via Structure-from-Motion (SfM) pipelines like COLMAP or GLOMAP.

In terms of data format, we use a structure similar to DSAC*, but with some adjustments for performance: rather than storing each image and pose as separate small files, we pack the images, intrinsics, and extrinsics into a few consolidated files. This avoids performance bottlenecks, especially on distributed file systems.

I'm currently working on documentation to describe the expected format in more detail, and I’m also considering adding a utility script to help convert outputs directly from COLMAP into the format used by SCRStudio. Stay tuned!

Let me know if you run into any specific issues or have further questions in the meantime — happy to help!

[xjiangan]

Hi! I’ve recently pushed a commit that documents the expected dataset structure and coordinate conventions used in SCRStudio. You can check it out in docs/data_conventions.md

Let me know if anything's unclear or if you run into issues using your dataset!