ROI Patches Plugin for FiftyOne

A FiftyOne plugin that tiles images into a configurable grid of ROI (Region of Interest) patches for region-based analysis.

Instead of saving physical cropped files, the plugin leverages FiftyOne's native to_patches() mechanism — it adds bounding box detections representing grid tiles to each sample. After creation, select the patches field via the Patches button to view the individual cropped tiles.

Installation

fiftyone plugins download https://github.com/mgustineli/roi-patches-plugin

Operators

Create ROI Patches

Tiles images into a grid of ROI patches.

• Rows / Columns — grid dimensions (1-20)
• Overlap % — overlap between adjacent tiles (0-90%)
• Field name — sample field to store the detections (default: roi_patches)

Supports delegated execution (schedule on builtin) for large datasets.

Clear ROI Patches

Removes an ROI patches field from the dataset. Select any existing Detections field from the dropdown.

Usage

1. Open the FiftyOne App and load a dataset
2. Open the operator search and look for "Create ROI patches"
3. Configure the grid and click Schedule
4. Once complete, click the Patches button in the App toolbar and select your field to view the individual tiles

To remove patches, open the operator search and look for "Clear ROI patches", then select the field to delete.

How It Works

The grid math computes bounding boxes in normalized [0, 1] coordinates:

• tile_w = 1 / (cols - (cols - 1) * overlap) for cols > 1, else 1.0
• tile_h = 1 / (rows - (rows - 1) * overlap) for rows > 1, else 1.0
• Each tile gets a label in R{row}_C{col} format (e.g., R0_C0, R2_C3)
• Tiles also store row, col, overlap_pct, grid_rows, and grid_cols as dynamic attributes