This document describes the CIFAR-10 dataset for classifying images into 10 categories. It contains 60,000 32x32 color images split into 50,000 training and 10,000 test images. Two methods are proposed: Method 1 extracts patches and features from each image and uses SVM/kNN, while Method 2 uses LoG and HoG features to preserve shape before SVM/kNN classification. Experiments test different parameters, with the best accuracy around 42% using a 13-dimensional Fisher vector and RBF SVM kernel.

1. CIFAR-10
Object Recognition in Images
Team Name: PatternfinderS Team # 24
Priyanshu Agrawal(201305511)
Satya Madala(201305508)

2. Background
 Image Classification
 Applications
Automatic image annotation
Reverse image search
 Kinds of datasets
Digital Images
Few thousands to millions of images.
 Machine Learning problems in this domain
Image search engine
Face detection

3. Problem
 Task is to classify images into following ten labels.
 The data labels are Cat, Dog, Airplane, Automobile,
Bird, Deer, Frog, Horse, Ship, Truck.
 Challenges
Large Data set
Small Image size (32x32)

4. Dataset
Data consists of 60,000 32x32 color images in 10
classes, with 6000 images per class.
There are 50,000 training images and 10,000 test images
The classes are completely mutually exclusive.
With the above mentioned samples, we will build a model
to identify whether the given image is Cat or a Dog etc.

5. Data Insights
 The objects within classes in this dataset are extremely
varied. For example the "bird" class contains many different
types of bird (both big birds and small).
 Not only are there many types of bird, but the occur at many
possible magnifications, all possible angles and all possible
poses.
 Sometimes only parts of the bird are shown.
 The poses problem is even worse for the dog/cat category,
because these animals occur at many different types of
poses, and sometimes only the head is shown or left part of
the body, etc.

6. Method 1
 Divide every 32*32 px image into 64 patches of 4*4 px
 Feature length would be 3*16*64 = 3072.
 Followed by Fisher on all images and extract their top 9
Dimensions
 Values obtained above will represent final feature
vector for images.
 Classification using SVM/kNN.

7. Method 1 (Continue…)
Block diagram 4
4
32
32
FDA
Patches
SVM/kNN
Patches
FDA
SVM /kNN Model
Labels
Train Data Test Data

8. Method 1 Drawbacks
 This method based on just intensity.
 Same object captured in different light.
 Image taken from different angles.

9. Method 2
Aim : Preserve the Shape of Image
Step 1 : Image is filtered using Laplace of Gaussian
filter(LoG).
log_image = edge(gray_image,'log', [], 2);

10. Method 2 (Continue…)
Step 2 : Feature is constructed by using histogram of
oriented gradients gives 324-d feature vector
hog_image= hog_feature_vector(log_image)

11. Method 2 (Continue…)
Block Diagram
FDA
LoG and HoG
SVM/kNN
LoG and HoG
FDA
SVM/kNN Model
Labels
Train Data Test Data

12. Experiments
 Divide given 50K data to train data and test data. Apply
Feature engineering and classification.
 40,000(80%) images as train data, 10,000(20%) images
as validation data.
 Result Evaluation : Percentage of labels that are
predicted correctly
 Number of Fisher dimensions (D = 7, 9, 13)
SVM kernel (linear, rbf)
kNN( k= 13, 17, 21)

13. Results
Benchmark : If we label all image as cat, accuracy is 10%
Classifier # Fisher
Dimentions
Method 1 Method 2
KNN (k=21) 7 20.59 25.65
KNN (k=21) 9 22.43 28.34
KNN (k=21) 13 22.68 28.71
SVM 7 30.98 40.45
SVM 9 34.56 42.85
SVM 13 34.83 -

14. Thank You