Computer Vision Projects

Three end-to-end computer vision projects, spanning classical image processing and modern deep learning. Each is a single self-contained, heavily documented notebook, delivered alongside a rendered artefact — a PDF write-up or an annotated video — so the results can be reviewed without running anything.

Project	What it does
face-detection-recognition	Finds faces in photographs and identifies who they belong to. Compares handcrafted descriptors (HOG), features learned from the data (PCA / eigenfaces) and deep face embeddings, then benchmarks a range of classifiers on top of each representation.
segmentation-classification-attacks	Understands natural scenes at two levels of granularity — which objects appear in an image, and which object each individual pixel belongs to — then attacks the resulting models with adversarial perturbations to probe how fragile they are.
object-detection	Tracks a puck and two strikers through air hockey footage using classical computer vision alone — colour thresholding, morphology, Hough circles and template matching — and renders the result as an annotated video that narrates each technique as it is applied.

Each folder has its own README with the full method breakdown.

Layout

face-detection-recognition/
    face-detection-recognition.ipynb    the project notebook
    face-detection-recognition.pdf      rendered write-up, 76 pages
    datasets/                           image data the notebook reads
segmentation-classification-attacks/
    segmentation-classification-attacks.ipynb
    segmentation-classification-attacks.pdf   rendered write-up, 116 pages
    clip_vitb16_v2_patch.pth            fine-tuned CLIP classifier head
object-detection/
    air-hockey-object-detection.ipynb   the processing pipeline
    video.mp4                           source footage
    processed-output.mp4                rendered result with subtitles
environment.yml                         conda environment for all three projects

Getting set up

conda env create -f environment.yml
conda activate cv

environment.yml covers all three projects. Its header documents an older, stricter pin set that the second project was originally validated against — worth reading before you rely on the exact numbers in that notebook.

A note on large files

clip_vitb16_v2_patch.pth (574 MB) is tracked with Git LFS, so you need git lfs install before cloning to get the real file rather than a pointer. The notebook loads it by name to skip retraining; delete it and the training path runs instead.

Both notebooks resolve their input data through a DATA_PATH constant defined near the top — point it at wherever your copy of the data lives.