DATA.md

Dataset and Pretrained Weight Download and Storage Layout

The current training and validation configs expect the following datasets:

Dataset / Weight	Used in	Download / Placement
Hypersim	training	Download Link
Real-World Benchmark	validation	Download Link
Synthetic Benchmark	validation	Download Link
DINOv3 vitl16 checkpoint	train infinidepth from scratch	place at ${commonspace}/pretrained_models/dinov3/dinov3_vitl16_pretrain_lvd1689m.pth

After downloading and unpacking them, place them under ${commonspace} like this:

${commonspace}/
├── datasets/
│   ├── Kitti/
│   ├── ETH3D/
│   ├── nyu/
│   ├── scannet/
│   ├── DIODE/
│   ├── cyberpunk/
│   ├── deadisland/
│   ├── spiderman2/
│   ├── spidermanmm/
│   └── watchdoglegion/
├── processed_datasets/
│   ├── hypersim/
│   │   └── train.txt
│   │   └── val.txt
│   ├── Kitti/
│   │   └── val.txt
│   ├── ETH3D/
│   │   └── val.txt
│   ├── nyu/
│   │   └── val.txt
│   ├── scannet/
│   │   └── val.txt
│   ├── DIODE/
│   │   └── val.txt
│   ├── cyberpunk/
│   │   └── val.txt
│   ├── deadisland/
│   │   └── val.txt
│   ├── spiderman2/
│   │   └── val.txt
│   ├── spidermanMM/
│   │   └── val.txt
│   └── watchdoglegion/
│       └── val.txt
└── pretrained_models/
    └── dinov3/
        └── dinov3_vitl16_pretrain_lvd1689m.pth

Current config paths

• data root: ${commonspace}/datasets/***
• meta file: ${commonspace}/processed_datasets/***/train.txt, ${commonspace}/processed_datasets/***/val.txt
• DINOv3 backbone weight: ${commonspace}/pretrained_models/dinov3/dinov3_vitl16_pretrain_lvd1689m.pth

Meta file format

Each line in a dataset meta file is interpreted relative to the dataset data_root and should be one of:

rgb_rel_path depth_rel_path
rgb_rel_path depth_rel_path prompt_depth_rel_path

Note

If you want to train/val on other datasets, you can prepare the data and meta file in the same format above, then modify the training/validation config to point to your new meta file and data root. You can also merge multiple meta files together and use --include to load them simultaneously.