Image Classification

Experimental benchmarking of classical computer vision descriptors and deep CNN features as inputs to an SVM classifier, evaluated on 5 selected classes from the Caltech-256 Object Categories dataset.

Dataset

Caltech-256 — 5 selected categories:

• backpack
• dog
• dolphin
• faces-easy
• soccer-ball

Set the dataset path via environment variable or edit src/utils/config.py:

export DATASET_PATH=/path/to/256_ObjectCategories

Feature Descriptors

Descriptor	Type	Module
HOG	Global	src/descriptors/hog_descriptor.py
LBP	Global	src/descriptors/lbp_descriptor.py
Gabor	Global	src/descriptors/gabor_descriptor.py
SIFT	Local	src/descriptors/sift_descriptor.py
VGG16 (middle layers)	Global / Local	src/descriptors/vgg_descriptor.py

Encoding Methods

Encoder	Module
Bag of Words (BoW)	src/encoders/bag_of_words.py
VLAD	src/encoders/vlad.py

Classifier

SVM with RBF kernel, 5-fold stratified cross-validation — src/classifiers/svm_classifier.py.

Project Structure

image-classification/
├── classification.ipynb     # Full pipeline notebook
├── src/
│   ├── descriptors/         # Feature extraction modules
│   ├── encoders/            # BoW and VLAD encoding
│   ├── classifiers/         # SVM + CV evaluation
│   └── utils/               # Config, data loader, visualization

Running

Open and run classification.ipynb end-to-end. Results (plots, JSON metrics) are saved to results/.