Datasets

Glue Factory supports several datasets for training and evaluation. Most datasets are implemented to download themselves automatically to the data/ directory.

Training Datasets

• Homographies (datasets.homographies):
  • Generates synthetic homographies from the Revisited Oxford-Paris 1M dataset.
  • Used for the first stage of training (pre-training).
• MegaDepth (datasets.megadepth):
  • Real-world outdoor scenes with depth and pose.
  • Used for fine-tuning.

Evaluation Benchmarks

• HPatches (datasets.hpatches): Planar scenes with illumination and viewpoint changes.
• MegaDepth-1500: A validation subset of MegaDepth for relative pose estimation.
• ScanNet-1500: Indoor scenes with ground truth pose.
• ETH3D: High-resolution images for line and point evaluation.

Helper Datasets

• datasets.image_folder: Loads images from a local directory without ground truth.
• datasets.image_pairs: Loads pairs of images defined in a text file.
• datasets.posed_images: Loads images with associated pose information.

Interface

All datasets inherit from BaseDataset. Key configuration parameters usually include:

• batch_size: Number of samples per batch.
• num_workers: Number of data loading threads.
• preprocessing: Image resizing and padding options (e.g., resize: 1024, side: long).
• grayscale: Convert images to grayscale on load.