Image Classification using the Spatial Pyramid Matching scheme in MATLAB

This project is inspired by the code snippets available from Svetlana Lazebnik et al.. It implement the Spatial Pyramid Matching scheme for classifying different scene categories, while yielding the GPU and Parallel Computing Toolboxes of MATLAB.

Progression of the loss function for optimizing some of the hyperparameters of the SVM

Progression of the loss function for optimizing one of the hyperparameters of the SVM

Code diagram of the script

Project Structure

• denseSIFTVN.m: Extracts dense SIFT descriptors from the images.
• DictionaryFormationVN.m: Forms a dictionary of visual words using k-means clustering.
• Final_Experiment.m: Main script to run the experiment, including feature extraction, dictionary formation, spatial pyramid matching, and SVM classification.
• gaussVN.m: Applies Gaussian filtering to the images.
• hist_intersection_VN.m: Computes the histogram intersection kernel.
• miniBatchKMeansVN.m: Performs mini-batch k-means clustering.
• resultsTable.mat: Stores the results of the experiments.
• scene_categories/: Directory containing the dataset of scene categories.
• SIFTnormalizationVN.m: Normalizes SIFT descriptors.
• SpatialPyramidVN.m: Constructs spatial pyramid representations of the images.
• splitTheDatastore2.m: Splits the image datastore into training and testing sets.

How to Run

To run this project:

1. Ensure MATLAB is installed on your system.
2. Clone this repository to your local machine.
3. Place your dataset in the scene_categories/directory.
4. Open MATLAB and navigate to the cloned project directory.
5. Run the Final_Experiment.m script to start the image classification pipeline.

run('Final_Experiment.m')

Disclaimer

This repository is a simple form of reproduction with a few changes compared to the initial files that are provided. In keeping with that theme, one can identify the use of GPU accelaration and Parallel processing to minimize the experiment's time to completion. Yet, there exist comments and parts inside some code snippets (even cases where the only changes made in the code snippets are just more explanatory comments) that belong in the initial draft of the contributors as the latter are provided in the link. All acknowledgements for those parts go to the authors!

License

This code is for teaching/research purposes only.

Table of results

Pyramid Levels	Number of Centers	Optimization Parameter	Mean Accuracy
2	200	BoxConstraint	71.7507
2	200	BoxConstraint & KernelScale	71.3947
2	200	All	71.6320
2	400	BoxConstraint	71.6024
2	400	BoxConstraint & KernelScale	71.8398
2	400	All	70.5935
3	200	BoxConstraint	76.0237
3	200	BoxConstraint & KernelScale	75.1335
3	200	All	76.0534
3	400	BoxConstraint	74.6588
3	400	BoxConstraint & KernelScale	75.9050
3	400	All	76.9139
4	200	BoxConstraint	73.2344
4	200	BoxConstraint & KernelScale	74.5994
4	200	All	75.5786
4	400	BoxConstraint	75.6083
4	400	BoxConstraint & KernelScale	75.0742
4	400	All	74.2433