traffic-image-classification

Traffic Image Classification

A data mining project that develops a predictive model that can determine, given images of traffic depicting different objects, which class (out of 11 total classes) it belongs to. The object classes are: car, suv, small_truck, medium_truck, large_truck, pedestrian, bus, van, people, bicycle, and motorcycle.

Exploratory Data Analysis (EDA)

• Features already extracted from images
• Features include:
  • 512 Histogram of Oriented Gradients (HOG) features
    • HOG counts occurrences of gradient orientation in localized portions of an image
  • 256 Normalized Color Histogram (Hist) features
    • Histogram gives intensity of distribution of an image
    • Get intuition about contrast, brightness, intensity distribution etc of that image
  • 64 Local Binary Pattern (LBP) features
    • LBP looks at points surrounding a central point and tests whether the surrounding points are greater than or less than the central point (i.e. gives a binary result).
    • Used for classifying textures (edges, corners, etc.)
  • 48 Color gradient (RGB) features
    • Color gradient measures gradual change/blend of color within an image
  • 7 Depth of Field (DF) features
    • Depth of Field is the distance about the plane of focus (POF) where objects appear acceptably sharp in an image
• Classes are imbalanced
  • 0 instances of human images
  • 3 instances of bicycle images
  • 10,375 instances of car images

Handling Class Imbalance

• Utilizing ML algorithms that can handle class imbalance by "balancing" the classes using weights
  • Weight minority classes higher than majority classes (using class_weight param)
  • Weighting is based on the class labels and is inversely proportional to the class frequencies in the input data

Data Preprocessing + Feature Selection

• StandardScaler to standardize the features (remove mean and scale to unit variance)
• Experiments with different feature selectors
• Measure performance of feature selection through cross validation of different models
  • PCA (Unsupervised feature selection)
    • Identify the combination of attributes (principal components) that account for the most variance in the data.
    • 95% kept variance = 14 components
    • 99% kept variance = 22 components
    • 100% kept variance = 34 components
  • LDA (Supervised feature extraction)
    • Identify attributes that account for the most variance between classes
    • 100% kept variance = 9 components
    • Did not perform well because of assumption that classes are normally distributed and have equal class covariance
  • Locally Linear Embedding (Non-linear dimensionality reduction)
    • Uses PCA so using 34 components
  • Removing features with low variance VarianceThreshold
    • Throw away features with 0 variance
      • remove the features that have the same value in all samples.
• Winner is VarianceThreshold

Model Selection & Tuning

• Use 5-Fold CV + Bayesian Optimization to individually tune each chosen model to to their own optimal hyperparameters
• Bayesian Optimization tries to approximate or fit a Gaussian process (a regression model known as a surrogate model) on function/model evaluations in order to propose sampling points in the search space of a model's hyperparameters using acquisition functions. Essentially, it is a smarter grid search, and empirically, it has been proven to be more efficient than a grid search and more effective than a random search.
• Models used in a VotingClassifier with soft voting:
  • Support Vector Machine (Gaussian RBF kernel)
  • Ensemble Methods:
    • Random Forest
    • Light Gradient Boosting (LightGBM)
    • Extra Trees Classifier
• VotingClassifier is then tuned using GridSearch + 5-Fold CV with different weights for each model
  • Optimal weights are determined to be each model with uniform weight (weight = 1)
• Model is evaluated using the validation set

Rank & F1 Score

My rank on the CLP public leaderboard is 1st, with a F1 score of 0.8598. This score is calculated on 50% of the test set.

Update: My rank on the CLP final leaderboard is 2nd with a F1 score of 0.8508. This score is calculated on the whole test set.

Note: This assignment follows a similar format with kaggle competitions in terms of ranking & scoring.