MICROSCOPY / SEGMENTATION BENCH
Pollenproof
Find the grains. Keep the dust out.
Calibrate a pollen counter
Recover all six grains with at least 90% mask overlap against the reference. Start with Otsu’s threshold suggestion, then tune the threshold and minimum area to retain faint pollen while discarding tiny dust. A low threshold finds more pixels; a high threshold loses dim grains.
Touching grains form one connected region. On the segmented view, tap the center of each grain in a touching cluster to place split markers. One marker alone does not split a cluster; two partition its pixels by nearest center. Then validate your count.
Colored outlines show accepted regions. Gold crosshairs are your split markers. Small bright specks are dust.
Image pipeline
Try the automatic threshold, then inspect faint grains and connected clusters.
Algorithm & source
Otsu threshold uses scikit-image’s cumulative class weights, means and between-class variance (BSD-2-Clause). Four-neighbor connected components, nearest-marker splitting, synthetic grayscale slides and ground-truth masks are original. This deliberately simplified pipeline is an educational calibration puzzle, not a clinical or environmental measurement tool.
Full source notices · scikit-image source