ahmedbrs/SketchMultiview_dataset

SketchMultiview Dataset

Multi-view image dataset for sketch-conditioned novel view synthesis.
Each sample contains 33 views of a scene at 480×480 resolution (4 elevation rings × 8 azimuths + original).

Dataset Structure

{split}/
├── generated_gt/          # Multi-view images (480×480 PNG)
│   ├── {group}/{coco_id}/ # Train: grouped into subfolders
│   │   ├── 00_original.png
│   │   ├── 00_azim000_elev+00.png
│   │   ├── 01_azim045_elev+00.png
│   │   ├── ...
│   │   ├── 31_azim315_elev-30.png
│   │   └── metadata.json
│   └── {coco_id}/         # Test/Val: flat structure
├── sketches/              # Sketch conditioning images (512×512 grayscale PNG)
├── captions/              # Text captions (.txt)
└── correspondences/       # COLMAP correspondences (JSON, 1024px coordinates)

Splits

Split	Samples	Size
Train	8,304	~102 GB
Test	477	~5.9 GB
Val	441	~5.5 GB

View Layout (33 views per sample)

• Elevation 0°: 8 views at 45° azimuth intervals + original
• Elevation +30°: 8 views at 45° azimuth intervals
• Elevation +60°: 8 views at 45° azimuth intervals
• Elevation -30°: 8 views at 45° azimuth intervals

Download & Extract

# Download all files
huggingface-cli download ahmedbrs/SketchMultiview_dataset --repo-type=dataset --local-dir .

# Extract test and val
tar xf test.tar
tar xf val.tar

# Reassemble and extract train (split into 45GB parts)
cat train.tar.part_aa train.tar.part_ab train.tar.part_ac | tar xf -

Correspondences

COLMAP-extracted correspondences are stored per sample in JSON format with pixel coordinates at 1024×1024 resolution. The training pipeline maps these to latent patch indices via:

patch = int(pixel_coord * latent_dim / 1024)

Usage with Training Pipeline

# Point to the extracted dataset directory
./train_ucpe.sh --fal-multiview  # set FAL_MULTIVIEW_DIR to extracted path