Project idea · ungraded

Simple Object Counter

Count objects in images — coins on a table or cell nuclei under a microscope — and report their sizes. Compare your counts with the ground truth and find out which images are hard and why.

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Load an image, convert to greyscale, blur and threshold with Otsu.
02Clean the mask with morphological opening (cv2.morphologyEx) to remove speckles.
03Count objects with connectedComponentsWithStats and filter by area.
04Draw numbered bounding boxes and save an annotated image.
05On a labelled dataset, compare counts with the ground truth and report the mean absolute error.

Stretch goals

Separate touching objects with the distance transform and watershed.
Match predicted boxes to ground-truth boxes with IoU and report precision and recall.
Measure object sizes in millimetres using a reference object of known size.

Data

BBBC039: fluorescent nuclei images with ground-truth masks (Broad Bioimage Benchmark Collection).
A coins image ships with scikit-image: skimage.data.coins() (data module), or photograph coins yourself.

Run it yourself

pip install opencv-python numpy scikit-image
python count.py image.png