From a95fc4c4050636ed5aa06e49842bc52666b1b7bf Mon Sep 17 00:00:00 2001 From: nunzip Date: Mon, 4 Feb 2019 17:30:26 +0000 Subject: Add Assignment Instructions --- report/paper.md | 40 ++++++++++++++++++++++++++++++++++++++-- 1 file changed, 38 insertions(+), 2 deletions(-) (limited to 'report') diff --git a/report/paper.md b/report/paper.md index 5f23d05..45d7eb2 100644 --- a/report/paper.md +++ b/report/paper.md @@ -1,6 +1,42 @@ -# Chapter I +# K-means codebook -## Sub-chapter I +We randomly select 100k descriptors for K-means clustering for building the visual vocabulary +(due to memory issue). Open the main_guideline.m and select/load the dataset. +>> [data_train, data_test] = getData('Caltech'); % Select dataset +Set 'showImg = 0' in getData.m if you want to stop displaying training and testing images. +Complete getData.m by writing your own lines of code to obtain the visual vocabulary and the +bag-of-words histograms for both training and testing data. Show, measure and +discuss the followings: + +## Vocabulary size + +## Bag-of-words histograms of example training/testing images + +## Vector quantisation process + +# RF classifier + +Train and test Random Forest using the training and testing data set in the form of bag-of-words +obtained in Q1. Change the RF parameters (including the number of trees, the depth of trees, the +degree of randomness parameter, the type of weak-learners: e.g. axis-aligned or two-pixel test), +and show and discuss the results: + +## recognition accuracy, confusion matrix, + +## example success/failures, + +## time-efficiency of training/testing, + +## impact of the vocabulary size on classification accuracy. + +# RF codebook + +In Q1, replace the K-means with the random forest codebook, i.e. applying RF to 128 dimensional +descriptor vectors with their image category labels, and using the RF leaves as the visual +vocabulary. With the bag-of-words representations of images obtained by the RF codebook, train +and test Random Forest classifier similar to Q2. Try different parameters of the RF codebook and +RF classifier, and show/discuss the results in comparison with the results of Q2, including the +vector quantisation complexity. # References -- cgit v1.2.3-54-g00ecf