BiliSakura/MACIV-T-2025

MAVIC-T Dataset

1. Refined Training Dataset (MACIV-T-2025-Structure-Refined)

Root: MACIV-T-2025-Structure-Refined
Layout: {task}/train/{input,target}/
Manifests: manifests/refined_manifest.csv, manifests/refined_manifest_crop_aug.csv

The refined dataset contains preprocessed and paired samples for supervised training. Augmented versions (*_crop_aug) are generated by scripts/prepare_refined_dataset_crop.py.

Spatial Alignment: SAR tasks (sar2ir, sar2rgb, and their augmented variants) use geographic reprojection to ensure strict pixel-level alignment between SAR scenes and the tile's IR/RGB mosaics. The SAR scene is reprojected into the target CRS, and both are cropped to their geographic overlap before resampling to 1024x1024.

Task Summary

Task	Resolution	Input Format	Target Format	Samples	Disk Size
sar2eo	256x256	1-band PNG	1-band PNG	89,411	7.0 GB
rgb2ir	1024x1024	3-band TIFF	1-band TIFF	2,272	8.9 GB
sar2ir	1024x1024	1-band TIFF	1-band TIFF	10,576	21 GB
sar2rgb	1024x1024	1-band TIFF	3-band TIFF	10,576	42 GB
rgb2ir_crop_aug	1024x1024	3-band TIFF	1-band TIFF	7,388	29 GB
sar2ir_crop_aug	1024x1024	1-band TIFF	1-band TIFF	74,317	146 GB
sar2rgb_crop_aug	1024x1024	1-band TIFF	3-band TIFF	75,703	297 GB
sar2rgb_sup	1024×1024	1-band PNG	3-band PNG	7,733	~17 GB

Crop-aug sets use stride=512 sliding windows over the georef-aligned overlap, producing dense overlapping crops for data augmentation.

Statistics of sar2rgb_sup

Supervised SAR→RGB pairs from external datasets. Paired list: manifests/paired_sar2rgb_sup.txt.

Source	Resolution	Input	Target	Samples	Disk Size	Notes
OpenEarthMap-SAR (SARLANG-1M OEMSAR)	1024×1024	1-band PNG	3-band PNG	4,813	~10 GB	OEMSARtrain + OEMSARtest, all as train
SpaceNet6 (SARLANG-1M)	900×var → 1024×1024	1-band PNG	3-band PNG	2,310	~5 GB	Variable native (e.g. 900×666); resized to 1024×1024
FUSAR-Map	1024×1024	1-band PNG	3-band PNG	610	~2 GB	rgb/ + sar/ TIF→PNG (no resize)

Sources with non-standard resolution (SpaceNet6) follow prepare_refined_dataset_crop.py practice: images smaller than 1024 are resized to 1024×1024 via bilinear resampling.

Technical Details

• Filenames: {city}__{tile}__{source}__{modality}.tif (crops append __cropXXX_xYY_yZZ).
• Normalization: Manhattan uint16 values are normalized to uint8 during preprocessing.
• Split: Entire refined set is for training; no internal val/test split.

---

2. Evaluation Inputs (Refined Root)

Used for validation and final testing. These folders contain inputs only (no targets) and now live alongside the refined training data.

• Validation: MACIV-T-2025-Structure-Refined/val/{rgb2ir,sar2ir,sar2rgb}
  • Files are located in val/{task}/input/ subdirectories
• Test: MACIV-T-2025-Structure-Refined/test/{rgb2ir,sar2ir,sar2rgb,sar2eo}
  • Files are located directly in test/{task}/ directories
  • Note: sar2eo only exists in test split, not val split. The dataset loader will automatically fall back to test split when val is requested for sar2eo.

---

3. Paired Validation Manifests (Golden Val Set)

Under manifests/, the following files define a small paired golden validation set with known input–target pairs, used for validation logging during training (e.g. logging reconstruction metrics on a fixed set of pairs):

Manifest	Task	Description
paired_val_sar2eo.txt	sar2eo	SAR → EO paired samples
paired_val_sar2ir.txt	sar2ir	SAR → IR paired samples
paired_val_sar2rgb.txt	sar2rgb	SAR → RGB paired samples
paired_val_rgb2ir.txt	rgb2ir	RGB → IR paired samples
paired_sar2rgb_sup.txt	sar2rgb_sup	Full SAR→RGB paired list (OpenEarthMap-SAR, SpaceNet6, FUSAR-Map; all train)

Format: Each line is input_path<TAB>target_path. paired_sar2rgb_sup.txt uses paths relative to the dataset root (portable across devices). paired_val_*.txt may use absolute paths; run scripts/rewrite_manifest_paths.py in 4th-MAVIC-T to rewrite for your setup.

Overlap with train: The paired val set is a subset of the paths in refined_manifest.csv / refined_manifest_crop_aug.csv. When loading the train set, the example trainers (DDBM, BiBBDM, I2SB, CUT, etc.) exclude all paths listed in paired_val_<task>.txt so that the training set and the golden val set do not overlap.

Generation: These manifests are produced by scripts/create_paired_validation_set.py, which selects samples from the refined training pool using MaRS embedding-based selection (centroid-anchored farthest-point strategy) to balance representativeness and diversity. By default, sar2eo gets 16×N pairs and the other tasks get N pairs (e.g. --n 4). Run from the repo root:

python scripts/create_paired_validation_set.py --dataset_root ./datasets/BiliSakura/MACIV-T-2025-Structure-Refined --n 4 --batch_size 8

Use these manifests for log validation: a fixed set of paired samples to compute and log metrics (e.g. PSNR, SSIM) during training.

---

4. Usage

Load data via the HuggingFace datasets wrapper:

from src.mavic_t_dataset import MavicTImageToImageDataset

loader = MavicTImageToImageDataset()
# Load base sar2ir training data
ds_train = loader.load(split="train", task="sar2ir")
# Load augmented rgb2ir training data
ds_aug = loader.load(split="train", task="rgb2ir_crop_aug")
# Load original validation inputs
ds_val = loader.load(split="val", task="sar2ir")
# For golden val set (log validation), read manifests/paired_val_<task>.txt (input<TAB>target per line)