How to finetune RoMa on a custom dataset with only image pairs and ground-truth match points (no depth data)?

[shahzeen031]

Hi, thank you for releasing this excellent work! 🙌

I’m trying to finetune RoMa on my own custom dataset, but my data is different from the MegaDepth setup.

My dataset contains image pairs from multiple modalities (e.g., RGB–thermal, RGB–NIR, etc.).

For each pair, I already have ground-truth pixel correspondences (matching points).

However, I do not have depth maps or camera intrinsics/extrinsics like in MegaDepth.

I saw in your documentation and code that the training pipeline in RoMa (and DKM) expects depth supervision and pose information (as used in MegaDepth and ScanNet).
In my case, I only have explicit match coordinates.

[liuyang3045]

The corresponding point pairs of feature points can be used to calculate the homography matrix or fundamental matrix of two images, thereby obtaining the pixel-level correspondence between the two images, which can replace the image transformation relationship verification based on depth maps in the source code.