| Image Restoration and Real-World Applications |
Super-resolution, denoising, deblurring and inpainting, then anomaly detection and defect inspection, medical imaging, remote sensing and transportation |
Open |
| Geometric and 3D Vision |
Camera geometry and calibration, two-view geometry, stereo, monocular depth, camera pose and structure from motion, SLAM, and feed-forward 3D models |
Open |
| Lab 1: Noise2Noise |
Train the Noise2Noise U-Net once with clean targets and once with only a second noisy photo as the target, and compare them with PSNR and SSIM |
Notebook Colab |
| Lab 2: Stereo Matching with a Learned Cost Volume |
On KITTI 2015, compare OpenCV's semi-global matching with a small GC-Net trained from scratch, then turn disparity into depth and a 3D point cloud |
Notebook Colab |
| Optional: Brain MRI Tumour Segmentation |
Write the Dataset class and train a U-Net with a pretrained encoder to segment tumours in LGG brain MRI scans, with every step worked through |
Notebook Colab |
| Optional: Fire and Smoke Segmentation with YOLO |
Segment fire and smoke with an Ultralytics YOLO segmentation model and inspect its masks against the labels |
Notebook Colab |
| Optional: 3D Point Cloud Classification with PointNet |
Classify ModelNet40 shapes directly from unordered point sets with PointNet and its T-Net alignment networks |
Notebook Colab |
