CSC3061 - Group Project

Project Sections

Section 1 - Training the Automatic System

• 32 Marks
• Two Main Components
  • A feature descriptor
    • Describe an image region with a high-dimensional descriptor
    • Options:
      1. Full Image (Prac 5 & 6)
      2. Dimensionality reduction techniques (Prac 7)
      3. Gabor Features
  • A learning method
    • Learn to classify an image region as a person or not.
    • Options
      1. SVM
      2. Nearest neighbour
      3. K-NN
• Folder images contains crops to be used for training.
• Justify your choice and the parameter values.

Section 2 - Testing the Classification System

• 21 Marks
• In order to justify previous choices, we need to divide dataset (both positive and negative examples) in 2 subsets: training and testing
  • Training samples and their labels for learning choose the best techniques, strategies and parameters for each block using the testing
  • Options:
    1. Given test and training files face_train.cdataset and face_test.cdataset
    2. Half/half
    3. Cross validation
• Evaluate the performance of your final choice(s)
  • Options:
    1. Recognition Rate
    2. TP, FP, TN, FN (Prac 7)
    3. Precision, recall, specificity, sensitivity, etc… (Prac 7)

Section 3 - Detection Implementation

• 37 Marks
• Apply your classification/verification system to implement a pedestrian detector
• Two important components
  • A sliding window detector. (Based on Prac 6)
    • Crop the image at every location and use the classifier to tell if that image region contains a face. By scanning every location on the full image, it will detect all instances of faces in that image.
    • In order to detect faces at multiple sizes, our sliding window detector should run at multiple scales (will require resizing image)
  • Non-maxima suppression. (Prac 6)
    • Overlapping detections are a common problem. NMS removes overlapping detection to improve performance. It keeps best detections in each region by selecting the strongest responses.
• Run your detector on the 4 files imX.jpg, and evaluate its performance
  • Calculate the performance of our detector (TP, FP, TN, FN, etc…)
  • Generate output images with bounding boxes
  • Reflect and explain the results that you have obtained
    • Why and when it fails?