Integrating AerialMegaDepth dataset

[kvuong2711]

Summary:

AerialMegaDepth (CVPR'25) is a hybrid varying-altitude 3D dataset combining MegaDepth images with geospatial mesh renderings, featuring 132K images across 137 scenes with camera intrinsics, poses, and depths in a unified coordinate frame. It has two potential benefits:

• Improve 3D estimation in aerial-ground scenarios (initial observation shown in paper: DUSt3R finetuned on AerialMegaDepth improves aerial-ground reconstruction)
• Metric scale: by co-registering MegaDepth images with geo-referenced mesh renderings, the provided depths + cameras are in metric scale, thus could potentially help improving metric scale estimation

What's included:

• Data processing scripts: downloading + conversion to WAI-formatted data + aggregation

• WAI-formatted dataloader

Not included (deferred to follow-ups):

• Pre-computed metadata (e.g., covisibility matrix + MoGe depthmaps)

• Selection of covisibility threshold for best performance during training/finetuning

Screenshots

See below some visualizations of the cameras + point clouds as sanity checks.

cc: @Nik-V9