CSV Windowed Image Dataset

CSVWindowedImageDataset is the image-only counterpart to CSVWindowedSegmentationDataset. It reads specific patches from large images based on coordinates (offsets) defined in a CSV file, without requiring masks.

When to use

Scenario	Recommended dataset
Specific patches pre-selected for unsupervised learning/inference	CSVWindowedImageDataset
Patches with masks pre-selected in a CSV	CSVWindowedSegmentationDataset
Standard image-only dataset from pre-cut tiles	ImageDataset

This dataset is ideal for tasks like:

• Self-supervised pre-training (e.g., Contrastive Learning) where you want to sample patches from specific regions of interest.
• High-throughput inference over specific regions defined by an external tool.

How it works

The dataset works exactly like CSVWindowedSegmentationDataset but skips mask loading. It uses rasterio windowed read to efficiently load only the pixels within the specified window.

CSV Structure

The CSV must contain the following columns:

Column	Description
image	Path to the original full-size image.
row_off	Vertical offset (line) where the patch starts.
col_off	Horizontal offset (column) where the patch starts.
patch_size	Width and height of the patch (pixels).

Example CSV:

image,row_off,col_off,patch_size
/data/scene_a.tif,0,0,256
/data/scene_a.tif,128,128,256
/data/scene_b.tif,1024,2048,256

YAML configuration

train_dataset:
  _target_: pytorch_segmentation_models_trainer.dataset_loader.image_dataset.CSVWindowedImageDataset
  input_csv_path: /data/unlabeled_patches.csv
  image_key: image
  row_off_key: row_off
  col_off_key: col_off
  patch_size_key: patch_size
  image_dtype: uint8
  augmentation_list:
    - _target_: albumentations.HorizontalFlip
      p: 0.5
    - _target_: albumentations.Normalize
      mean: [0.485, 0.456, 0.406]
      std: [0.229, 0.224, 0.225]
    - _target_: albumentations.pytorch.ToTensorV2
  data_loader:
    batch_size: 32
    num_workers: 4
    shuffle: true

Comparison with ImageDataset

Property	CSVWindowedImageDataset	ImageDataset
I/O Strategy	Windowed read (rasterio)	Full file read (PIL/rasterio)
Image Source	Large rasters	Pre-cut tiles
Memory usage	Low (independent of raster size)	High (proportional to file size)
Flexibility	High (on-the-fly crops)	Fixed tiles