Machine Learning - Object Detection and Classification

CS771A Project Presentation
Machine Learning - Tools, Techniques and Applications
Group 43
Preetansh Goyal, Shubham Gupta, Vandana Gautam, Vikas Jain
IIT KANPUR
India
10-4-2016
Group 43 Preetansh Goyal, Shubham Gupta, Vandana Gautam, Vikas Jain (IIT Kanpur)CS771A Project Presentation 10-4-2016 1 / 18

Table of Contents
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization
3 Experiments and Results
4 Scope of Improvements

Outline
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization

Introduction
Aim Detection and classiﬁcation of objects in surveillance
video into various categories.
Categories - Car, Person, Motorcycle, Bicycle,
Rickshaw, Autorickshaw
Figure: Detection and Classiﬁcation of objects
Motivation - Recognition, Tracking, Security, etc.

Outline
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization

Feature Extraction - Dataset
To obtain the image dataset required for classiﬁcation, we extracted
frames from the provided video data.
Frames were taken after every 20 frames of each video.
The images in the dataset correspond to the labelled bounding boxes
from these frames.
Figure: Red Boxes Images used as dataset

Feature Extraction - Feature Vector
Bag of visual words[1]
Computed ∼ 300 SIFT points per image in the dataset.
Used mini-batch K-means - 700 clusters of siftpoints.
Voting of siftpoints of each image in clusters give vector of dimension
700.

700.
HOG features[2]
Images in the dataset are resized to 128x128.
Obtained HOG feature vector of dimension 15876 for each image.

700.
HOG features[2]
Images in the dataset are resized to 128x128.
Obtained HOG feature vector of dimension 15876 for each image.
Convolution Neural Network features
Used VGG 16 Model[3].
Output of FC-4096 layer as feature vector.

Outline
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization

Classification - Classifiers Used
After obtaing feature vectors of images in the dataset, classifiers are
trained given the category of each image.
Classifiers Used:
Support Vector Machines
Linear Kernel
Gaussian Kernel
Random Forest Classifier
Decision Tree

Outline
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization

Object localization
Objects of interest are extracted from the test videos. Three techniques
are implemented for the task.

Object localization
Using Background Subtraction
Using Mixture Of Gaussions(MOG)

Object localization
Measuring Optical Flow

Object localization
Sliding Window Method

Object localization
Figure: Contours are drawn on right image

Object localization
Figure: Contours are drawn on right image
Then, Learned Classiﬁer is used to classify extracted image.

Outline
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization

Experiments and Results
Dataset for classiﬁer learning
Labelled dataset was noisy - separated good images for dataset
creation.
∼ 1600 training images and ∼ 450 test images.
Around 300 training images of each class used for training.

Experiments and Results
Dataset for classifier learning
Labelled dataset was noisy - separated good images for dataset
creation.
∼ 1600 training images and ∼ 450 test images.
Around 300 training images of each class used for training.
Accuracy Obtained
Classification Accuracies Obtained
Classifier/features BOVW HOG CNN
SVM(linear) 0.8535 0.9129 0.9873
SVM(gaussian) 0.879 0.756 0.9809
Random Forest 0.8322 0.9023 0.9745
Decision Tree 0.6475 0.6072 0.8917

Outline
1 Introduction
2 Methodology
Feature Extraction
Classiﬁcation
Object localization

Scope of Improvements
Dataset Dataset and Labels are noisy, improvement of dataset are
likely to give improve results.

Localization BGS, Optical ﬂow, Sliding window do not work good for
localization. Learning method can be used for localization.

Localization BGS, Optical flow, Sliding window do not work good for
localization. Learning method can be used for localization.
CNN Though the accuracy of classifier with CNN features is very
good, it require 1-2 second each frame for the classification.
For real time classification, Fast R-CNN[4] can be used.

References I
Object recognition from local scale-invariant features
Lowe, David G. (1999).
Proceedings of the International Conference on Computer Vision. pp.
11501157
Histograms of oriented gradients for human detection
Dalal, Navneet, and Bill Triggs.
Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE
Computer Society Conference on. Vol. 1. IEEE, 2005
Very Deep Convolutional Networks for Large-Scale Image Recognition
K. Simonyan, A. Zisserman.
arXiv:1409.1556

References II
Fast R-CNN
Ross B. Girshick.
Proceedings of the IEEE International Conference on Computer
Vision. 2015.
Libraries Used
OpenCV(i/o, bgs, optical ﬂow, sift), sklearn(k-means, classiﬁers),
sklearn-image(hog).

Machine Learning - Object Detection and Classification

More Related Content

What's hot

Viewers also liked

Similar to Machine Learning - Object Detection and Classification

Recently uploaded

Machine Learning - Object Detection and Classification