CS 766 Assignment 3: Locality-constrained Linear Coding for Scene Classification

Contents

1. Introduction
2. Hard Code Word
1. Results
3. Locality-constrained Linear
1. Results
4. Grid Search
5. Sequential Hierarchy Classifier
1. Manually assigned clusters
1. Results
2. Clusters from K-means
1. Results
6. Other Dataset Evaluation
1. Birds
2. Butterflies
7. Other Experiments
1. Results
8. Scene Datasets
9. Code
10. Git Logs
11. References

Locality-constrained Linear Coding (LLC)

1. Rather than assigning each SIFT descriptor to just 1 of the M clusters, find the k-nearest neighbors (out of the M clusters) for each of the SIFT descriptors in each image. We chose to use k = 5.


2. Then, use those nearest neighbors to reconstruct the feature x to be an M x 1 vector where that vector has k non zero values. These k values will be normalized according to the relevant clusters' distances to the particular SIFT descriptor.


3. Rather than using a sum pooling method and concatenating all of the features for each sub region in the spatial pyramid scheme, LLC uses a max pooling method. This involves taking the maximum c value (cluster assignment value) to construct the c vector for each pyramid level, and then normalizing it by the length of that c vector.