blackFLINT/CLASSICAL-OBJECT-DETECTION-AND-LOCALISATION-SYSTEM-

This project develops a classical (non-deep-learning) computer vision system that detects and localises the University of Ghana logo in still images using OpenCV # Classical Object Detection and Localisation System Target object: University of Ghana logo (procedural placeholder crest — see note below) ## Quick start ## python data_generator.py builds data/ (70 test images + template + annotations.json) ```bash pip install opencv-python numpy matplotlib python experiments.py # runs every experiment in the report, writes results/plots/ ``` ## Layout ``` data_generator.py dataset + ground-truth generation detector/ utils.py IoU, non-max suppression masked_ncc.py masked mean-subtracted NCC (see docstring for derivation) template_matcher.py multi-scale masked-NCC detector sliding_window.py window/step search, NumPy-scored feature_matcher.py ORB + RANSAC homography detector evaluation.py IoU matching -> Precision / Recall / F1 visualize.py box drawing, image-grid panels experiments.py every experiment + plot used in the report data/ generated dataset (templates/, test/images/, annotations.json) results/plots/ all figures and JSON result dumps referenced in the report ``` ## Two real bugs fixed during development (documented in the report, Sections 3.1 & 5.1–5.2) - An early template with unmasked black corners produced strongly *negative* correlation at the true object location because test backgrounds weren't black. Fixed with alpha-channel masking. - OpenCV's only built-in masked correlation mode (`TM_CCORR_NORMED`) is not mean-subtracted and scored ~0.86 almost everywhere regardless of structure. Fixed by deriving and implementing a proper masked, mean-centred NCC from scratch (`detector/masked_ncc.py`).