Computer Vision:
Harris, SIFT & YOLO

Overview

A comprehensive computer vision assignment implementing five foundational algorithms: Harris Corner Detection, Image Pyramids, ORB feature detection, SIFT keypoint detection, and YOLOv5 real-time object detection on the COCO dataset.

Q1 — Harris Corner Detection

The Harris algorithm detects corners by analysing local intensity changes in multiple directions. Applied to a Rubik's Cube image — a geometrically rich subject with many well-defined corners. All 27 visible faces and edge intersections were correctly detected.

corners = cv2.cornerHarris(image, blockSize=2, ksize=3, k=0.04)
corners = cv2.dilate(corners, None)
image_with_corners[corners > 0.01 * corners.max()] = [0, 0, 255]

Q2 — Image Pyramid

Gaussian image pyramid constructed over 6 downsampling levels using cv2.pyrDown(). Applied to a building photograph to demonstrate multi-scale image representation — foundational for scale-invariant detection and image blending.

Q3 — ORB Feature Detection

SURF is patent-restricted in OpenCV, so ORB (Oriented FAST and Rotated BRIEF) was used as a free, high-performance alternative. ORB detects keypoints and computes binary descriptors suitable for fast matching.

orb = cv2.ORB_create()
keypoints, descriptors = orb.detectAndCompute(gray, None)
img_with_keypoints = cv2.drawKeypoints(
    image, keypoints, None, (255, 0, 0),
    flags=cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS
)

Q4 — SIFT Keypoint Detection

SIFT (Scale-Invariant Feature Transform) detects and describes local features invariant to scale, rotation, and illumination changes. The rich circular keypoint visualisation shows scale and orientation for each detected feature.

sift = cv2.SIFT_create()
keypoints, descriptors = sift.detectAndCompute(gray, None)
img_with_keypoints = cv2.drawKeypoints(image, keypoints, None)

Q5 — YOLOv5 Object Detection (COCO Dataset)

YOLOv5 (via Ultralytics) was loaded with pre-trained weights and run on 5 test images from the COCO dataset. YOLO performs single-pass detection — classifying and localising all objects in one forward pass.

from ultralytics import YOLO
model   = YOLO("yolov5s.pt")
results = model(img_rgb)
boxes   = results[0].boxes
# boxes.cls   → class IDs
# boxes.xyxy  → bounding box coords
# boxes.conf  → confidence scores

YOLO Detection Results

• Image 1 — 1 person, 1 skateboard detected (conf: 0.82, 0.71)
• Image 2 — 3 persons, 3 elephants detected (conf up to 0.93)
• Image 3 — 1 apple, 1 orange detected (conf: 0.85, 0.52)
• Image 4 — 2 persons, 5 cows detected (conf up to 0.91)
• Image 5 — 2 airplanes detected (conf: 0.90, 0.89)

Tech Stack

• opencv-python 4.10 — Harris, pyramids, ORB, SIFT
• ultralytics 8.3 — YOLOv5 inference
• torch 2.4 + torchvision — deep learning backend
• COCO dataset — 80 object classes, industry-standard benchmark
• Google Colab — execution environment