FlowROI

ROI based compression algorithm for ComplexEye array microscope

Code will be released upon paper publication.

FlowROI: Optical-Flow-Based ROI Extraction

FlowROI extracts motion-based regions of interest (ROIs) from image sequences. It estimates optical flow between neighboring frames, converts the flow into a motion saliency map, thresholds the most salient pixels, and cleans the result with morphology and connected-component filtering.

Main features

• CPU backends: DIS optical flow and Farnebäck optical flow
• Optional CUDA Farnebäck / TV-L1 backends if OpenCV was built with CUDA
• Translation-only stabilization using phase correlation
• Fast histogram-based percentile thresholding
• Morphological cleanup and small-component removal
• Optional debug outputs for registered frames, saliency maps, and raw masks

Installation

pip install opencv-python numpy

For CUDA backends, install an OpenCV build with CUDA support.

Usage

python flow_roi_extractor.py \
  --input-dir ./dataset \
  --output-dir ./flow_output \
  --pattern "*_data.bmp" \
  --step 1 \
  --backend dis \
  --keep-top 0.20

The output directory will contain one *_roi.png binary mask for each input frame.

Useful arguments

• --backend: dis, farneback, cuda_farneback, or cuda_tvl1
• --keep-top: fraction of highest-saliency pixels kept as foreground, e.g. 0.20
• --step: temporal gap between paired frames
• --min-area-bin: minimum component size after initial thresholding
• --min-area-clean: minimum component size in final mask
• --save-debug: save intermediate images