Project 3: Scene Classification — Sam Birch (sbirch)

CS143 10/24

This project used a set of one vs. all SVMs to do scene classification on histograms of visual words (clustered from a random subset of dense SIFT features over the training set). The overall results were as expected, but the inter-class variation was very large, ranging from 10% to over 90%. Some of these misclassifications make a lot of sense (e.g. industrial/highway), while others are stranger (kitchen/forest). This is probably the artifact nonetheless of the particular vocabulary & data.

accuracy =

    0.6213