Document Comprehensive Training Data Setup and Processing

[engichang1467]

Detail Data Pipeline for Diverse Training Datasets

Description:

The original Rig3R model was trained on an extremely diverse mix of datasets, encompassing indoor, driving, synthetic, and object-centric scenes from sources including CO3D-v2, BlendedMVS, Map-free, ScanNet++ v2, MVImgNet, PointOdyssey, Virtual KITTI2, TartanAir V2, PandaSet, KITTI, Argoverse2, nuScenes, and Waymo.

For the repository to be fully reproducible, the documentation must detail how to assemble and process this extensive data mix, specifically:

1. Data Acquisition Guidance: Provide links or guidance for gathering the necessary external datasets.
2. Sampling Strategy: Detail the specific sampling procedures: how images from COLMAP datasets were sampled based on covisibility, the use of random stride for other datasets, and the method for subsampling rig cameras while ensuring the front-facing camera is included in rig-based datasets.
3. Sequence Configuration: Clarify how the training samples were configured (e.g., 24-frame samples).