CS 180 Computer Vision Projects

Project 00 / Camera geometry

Perspective & dolly zoom

I photographed close and distant views to study how camera position and focal length change apparent perspective, then assembled a dolly-zoom sequence.

Compared facial distortion and architectural perspective across shooting distances.
Kept a subject roughly fixed in frame while changing camera distance and zoom.

Animated dolly zoom from the camera geometry project — Dolly-zoom sequence from Project 0.

Project 01 / Image alignment

Colorizing glass-plate images

I aligned three vertically stacked monochrome exposures into one color image.

Built a coarse-to-fine image pyramid and used Sobel edge maps for alignment.
Searched channel shifts with normalized correlation, then cropped border artifacts.

Colorized historic cathedral scene after aligning three image channels — Aligned RGB reconstruction from the notebook output.

Project 02 / Spatial and frequency domains

Filters, hybrid images & blending

I implemented convolution and several multiscale image-processing techniques to test how structure changes across spatial frequencies.

Compared a from-scratch convolution with SciPy and explored finite differences, derivative-of-Gaussian filters, and sharpening.
Built hybrid images, Gaussian and Laplacian stacks, and multiresolution blends.

A hybrid cat image made by combining low and high spatial frequencies — One hybrid-image result; its appearance changes with viewing distance.

Project 03 / Registration

Manual & automatic photo mosaics

I first stitched photographs with manually selected correspondences, then automated the correspondence-finding stage.

Estimated homographies and warped images into a shared canvas.
Used Harris corners, adaptive non-maximal suppression, normalized patch descriptors, ratio-test matching, and RANSAC.
Compared automatic mosaics with manual-point results on three image pairs.

Automatic and manual panorama results shown side by side — Automatic result at left, manual correspondences at right. The visible distortions show where alignment and blending still need work.

Project 04 / Neural rendering

NeRF from real photographs

I trained a PyTorch neural radiance field on a small model car, using photos rather than only the provided synthetic scene.

Calibrated the camera, estimated poses from ArUco markers with solvePnP, and prepared 45 usable views: 36 for training and nine held out.
Implemented rays, positional encoding, point sampling, a radiance-field MLP, and volume rendering.
Tracked held-out PSNR and rendered a novel camera path. The real-scene reconstruction remained visibly rougher than the synthetic result.

Novel camera path rendered by the NeRF model around a small model car — Rendered camera path from the real-scene model.

A reconstructed view of the model car with an ArUco marker visible — A single reconstructed view.

Project 05 / Generative models

Denoising & flow matching

I implemented and trained PyTorch UNets for image denoising and time- and class-conditioned flow matching on MNIST.

Inspected training curves, denoising at noise levels outside training, and generated samples.
Used a pretrained diffusion model in a separate notebook for iterative denoising and image-editing experiments.

Grid of MNIST digits sampled from a class-conditioned flow model — Class-conditioned samples from the saved notebook output.