Image Slicer¶

v2¶

Split input images into overlapping tiles or slices using the Slicing Adaptive Inference (SAHI) technique to enable small object detection by processing smaller image regions where objects appear larger relative to the image size, improving detection accuracy for small objects in large images through tiled inference workflows with equal-sized slices and deduplication.

How This Block Works¶

This block implements the first step of the SAHI (Slicing Adaptive Inference) technique by dividing large images into smaller overlapping tiles. This approach helps detect small objects that might be missed when processing the entire image at once. The block:

1. Receives an input image and slicing configuration:
2. Takes an input image to be sliced
3. Receives slice dimensions (width and height in pixels)
4. Receives overlap ratios for width and height (controls overlap between adjacent slices)
5. Calculates slice positions:
6. Generates a grid of slice coordinates across the image
7. Positions slices with specified overlap between consecutive slices
8. Overlap helps ensure objects at slice boundaries are not missed
9. Adjusts border slice positions to ensure all slices are equal size (pushes border slices toward image center)
10. Creates image slices:
11. Extracts each slice from the original image using calculated coordinates
12. Creates WorkflowImageData objects for each slice with crop metadata
13. Stores offset information (x, y coordinates) for each slice relative to original image
14. Maintains parent image reference for coordinate mapping
15. Deduplicates slices:
16. Removes any duplicate slice coordinates that may occur from overlap calculations
17. Ensures each unique slice position appears only once in the output
18. Prevents redundant processing of identical image regions
19. Handles edge cases:
20. Filters out empty slices (if any occur)
21. Ensures all slices fit within image boundaries
22. Creates crop identifiers for tracking each slice
23. Returns list of slices:
24. Outputs all unique slices as a list of images
25. All slices have equal dimensions (border slices adjusted to match)
26. Increases dimensionality by 1 (one image becomes multiple slices)
27. Each slice can be processed independently by downstream blocks

The SAHI technique works by making small objects appear larger relative to the slice size. When an object is only a few pixels in a large image, scaling the image down to model input size makes the object too small to detect. By slicing the image and processing each slice separately, the same object occupies more pixels in each slice, making detection more reliable. Overlapping slices ensure objects near slice boundaries are detected in at least one slice.

Common Use Cases¶

• Small Object Detection: Detect small objects in large images using SAHI technique (e.g., detect small vehicles in aerial images, find license plates in wide-angle camera views, detect insects in high-resolution photos), enabling small object detection workflows
• High-Resolution Image Processing: Process high-resolution images by slicing them into manageable pieces (e.g., process satellite imagery, analyze medical imaging scans, process large document images), enabling high-resolution processing workflows
• Aerial and Drone Imagery: Detect objects in aerial photography where objects are small relative to image size (e.g., detect vehicles in drone footage, find people in aerial surveillance, detect structures in satellite images), enabling aerial detection workflows
• Wide-Angle Camera Monitoring: Improve detection in wide-angle camera views where objects appear small (e.g., monitor large parking lots, detect objects in panoramic views, analyze traffic in wide camera coverage), enabling wide-angle monitoring workflows
• Medical Imaging Analysis: Analyze medical images by processing regions separately (e.g., detect lesions in large scans, find anomalies in medical images, analyze radiology images), enabling medical imaging workflows
• Document and Text Processing: Process large documents by slicing into regions (e.g., OCR large documents, detect text regions in scanned documents, analyze document layouts), enabling document processing workflows

Connecting to Other Blocks¶

This block receives images and produces image slices:

• After image input or preprocessing blocks to slice images for SAHI processing (e.g., slice input images, process preprocessed images, slice transformed images), enabling image-to-slice workflows
• Before detection model blocks (Object Detection Model, Instance Segmentation Model) to process slices for small object detection (e.g., detect objects in slices, run detection on each slice, process slices with models), enabling slice-to-detection workflows
• Before Detections Stitch block (required after detection models) to merge detections from slices back to original image coordinates (e.g., merge slice detections, combine detection results, reconstruct full-image predictions), enabling slice-detection-stitch workflows
• In SAHI workflows following the pattern: Image Slicer → Detection Model → Detections Stitch to implement complete SAHI technique for small object detection
• Before filtering or analytics blocks to process slice-level results before stitching (e.g., filter detections per slice, analyze slice results, process slice outputs), enabling slice-to-analysis workflows
• As part of multi-stage detection pipelines where slices are processed independently and results are combined (e.g., multi-scale detection, hierarchical detection, parallel slice processing), enabling multi-stage detection workflows

Version Differences¶

This version (v2) includes the following enhancements over v1:

• Equal-Sized Slices: All slices generated by the slicer have equal dimensions. Border slices that would normally be smaller in v1 are adjusted by pushing them toward the image center, ensuring consistent slice sizes. This provides more predictable processing behavior and ensures all slices are processed with the same dimensions, which can be important for model inference consistency.
• Deduplication: Duplicate slice coordinates are automatically removed, ensuring each unique slice position appears only once in the output. This prevents redundant processing of identical image regions that could occur due to overlap calculations, improving efficiency and preventing duplicate detections.

Requirements¶

This block requires an input image. The slice dimensions (width and height) should ideally match the model's input size for optimal performance. If slice size differs from model input size, slices will be resized during inference which may affect accuracy. Default slice size is 640x640 pixels, but this should be adjusted based on your model's input size (e.g., use 320x320 for models with 320 input size, 1280x1280 for models with 1280 input size). Overlap ratios (default 0.2 or 20%) help ensure objects at slice boundaries are detected, but higher overlap increases processing time. The block should be used with object detection or instance segmentation models, followed by Detections Stitch block to merge results. For more information on SAHI technique, see: https://ieeexplore.ieee.org/document/9897990. For a practical guide, visit: https://blog.roboflow.com/how-to-use-sahi-to-detect-small-objects/.

Type identifier¶

Use the following identifier in step "type" field: roboflow_core/image_slicer@v2to add the block as as step in your workflow.

Properties¶

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Available Connections¶

Input and Output Bindings¶

The available connections depend on its binding kinds. Check what binding kinds Image Slicer in version v2 has.

{
    "name": "<your_step_name_here>",
    "type": "roboflow_core/image_slicer@v2",
    "image": "$inputs.image",
    "slice_width": 320,
    "slice_height": 320,
    "overlap_ratio_width": 0.1,
    "overlap_ratio_height": 0.1
}

v1¶

Split input images into overlapping tiles or slices using the Slicing Adaptive Inference (SAHI) technique to enable small object detection by processing smaller image regions where objects appear larger relative to the image size, improving detection accuracy for small objects in large images through tiled inference workflows.

How This Block Works¶

This block implements the first step of the SAHI (Slicing Adaptive Inference) technique by dividing large images into smaller overlapping tiles. This approach helps detect small objects that might be missed when processing the entire image at once. The block:

1. Receives an input image and slicing configuration:
2. Takes an input image to be sliced
3. Receives slice dimensions (width and height in pixels)
4. Receives overlap ratios for width and height (controls overlap between adjacent slices)
5. Calculates slice positions:
6. Generates a grid of slice coordinates across the image
7. Positions slices with specified overlap between consecutive slices
8. Overlap helps ensure objects at slice boundaries are not missed
9. Border slices may be smaller than specified size to fit within image bounds
10. Creates image slices:
11. Extracts each slice from the original image using calculated coordinates
12. Creates WorkflowImageData objects for each slice with crop metadata
13. Stores offset information (x, y coordinates) for each slice relative to original image
14. Maintains parent image reference for coordinate mapping
15. Handles edge cases:
16. Filters out empty slices (if any occur)
17. Ensures all slices fit within image boundaries
18. Creates crop identifiers for tracking each slice
19. Returns list of slices:
20. Outputs all slices as a list of images
21. Increases dimensionality by 1 (one image becomes multiple slices)
22. Each slice can be processed independently by downstream blocks

The SAHI technique works by making small objects appear larger relative to the slice size. When an object is only a few pixels in a large image, scaling the image down to model input size makes the object too small to detect. By slicing the image and processing each slice separately, the same object occupies more pixels in each slice, making detection more reliable. Overlapping slices ensure objects near slice boundaries are detected in at least one slice.

Common Use Cases¶

• Small Object Detection: Detect small objects in large images using SAHI technique (e.g., detect small vehicles in aerial images, find license plates in wide-angle camera views, detect insects in high-resolution photos), enabling small object detection workflows
• High-Resolution Image Processing: Process high-resolution images by slicing them into manageable pieces (e.g., process satellite imagery, analyze medical imaging scans, process large document images), enabling high-resolution processing workflows
• Aerial and Drone Imagery: Detect objects in aerial photography where objects are small relative to image size (e.g., detect vehicles in drone footage, find people in aerial surveillance, detect structures in satellite images), enabling aerial detection workflows
• Wide-Angle Camera Monitoring: Improve detection in wide-angle camera views where objects appear small (e.g., monitor large parking lots, detect objects in panoramic views, analyze traffic in wide camera coverage), enabling wide-angle monitoring workflows
• Medical Imaging Analysis: Analyze medical images by processing regions separately (e.g., detect lesions in large scans, find anomalies in medical images, analyze radiology images), enabling medical imaging workflows
• Document and Text Processing: Process large documents by slicing into regions (e.g., OCR large documents, detect text regions in scanned documents, analyze document layouts), enabling document processing workflows

Connecting to Other Blocks¶

This block receives images and produces image slices:

• After image input or preprocessing blocks to slice images for SAHI processing (e.g., slice input images, process preprocessed images, slice transformed images), enabling image-to-slice workflows
• Before detection model blocks (Object Detection Model, Instance Segmentation Model) to process slices for small object detection (e.g., detect objects in slices, run detection on each slice, process slices with models), enabling slice-to-detection workflows
• Before Detections Stitch block (required after detection models) to merge detections from slices back to original image coordinates (e.g., merge slice detections, combine detection results, reconstruct full-image predictions), enabling slice-detection-stitch workflows
• In SAHI workflows following the pattern: Image Slicer → Detection Model → Detections Stitch to implement complete SAHI technique for small object detection
• Before filtering or analytics blocks to process slice-level results before stitching (e.g., filter detections per slice, analyze slice results, process slice outputs), enabling slice-to-analysis workflows
• As part of multi-stage detection pipelines where slices are processed independently and results are combined (e.g., multi-scale detection, hierarchical detection, parallel slice processing), enabling multi-stage detection workflows

Requirements¶

This block requires an input image. The slice dimensions (width and height) should ideally match the model's input size for optimal performance. If slice size differs from model input size, slices will be resized during inference which may affect accuracy. Default slice size is 640x640 pixels, but this should be adjusted based on your model's input size (e.g., use 320x320 for models with 320 input size, 1280x1280 for models with 1280 input size). Overlap ratios (default 0.2 or 20%) help ensure objects at slice boundaries are detected, but higher overlap increases processing time. The block should be used with object detection or instance segmentation models, followed by Detections Stitch block to merge results. For more information on SAHI technique, see: https://ieeexplore.ieee.org/document/9897990. For a practical guide, visit: https://blog.roboflow.com/how-to-use-sahi-to-detect-small-objects/.

Type identifier¶

Use the following identifier in step "type" field: roboflow_core/image_slicer@v1to add the block as as step in your workflow.

Properties¶

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Available Connections¶

Input and Output Bindings¶

The available connections depend on its binding kinds. Check what binding kinds Image Slicer in version v1 has.

{
    "name": "<your_step_name_here>",
    "type": "roboflow_core/image_slicer@v1",
    "image": "$inputs.image",
    "slice_width": 320,
    "slice_height": 320,
    "overlap_ratio_width": 0.1,
    "overlap_ratio_height": 0.1
}