Digital Image Analysis:
Image Pattern Classification
Image Classification: A core task in Computer Vision
Slide Credit: Fei Fei Li et al., Stanford
The Problem: Semantic Gap
Slide Credit: Fei Fei Li et al., Stanford
Challenges: Viewpoint variation
Slide Credit: Fei Fei Li et al., Stanford
Challenges: Illumination
Slide Credit: Fei Fei Li et al., Stanford
Challenges: Deformation
Slide Credit: Fei Fei Li et al., Stanford
Challenges: Occlusion
Slide Credit: Fei Fei Li et al., Stanford
Challenges: Background Clutter
Slide Credit: Fei Fei Li et al., Stanford
Challenges: Intraclass variation
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Brief History of Image Representation and Classification
Slide Credit: Fei Fei Li et al., Stanford
Convolutional Neural Networks (CNN) have
become an important tool for object
recognition
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei-Fei Li et al., Stanford
ImageNet Challenge: Classification Task
Slide Credit: Ava Soleimany, MIT
ImageNet Challenge: Classification Task
Slide Credit: Ava Soleimany, MIT
Convolutional Neural Networks (CNN)
were not invented overnight
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
CNN Architectures
for Image
Classification
CNN Architectures
Slide Credit: Fei Fei Li et al., Stanford
Review: LeNet-5
Slide Credit: Fei Fei Li et al., Stanford
Case Study: AlexNet
[Krizhevsky et al. 2012]
Slide Credit: Fei Fei Li et al., Stanford
Case Study: AlexNet
[Krizhevsky et al. 2012]
Slide Credit: Fei Fei Li et al., Stanford
ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) winners
Slide Credit: Fei Fei Li et al., Stanford
ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) winners
Slide Credit: Fei Fei Li et al., Stanford
ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) winners
Slide Credit: Fei Fei Li et al., Stanford
ZFNet
Slide Credit: Fei Fei Li et al., Stanford
ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) winners
Slide Credit: Fei Fei Li et al., Stanford
Case Study: VGGNet
[Simonyan and Zisserman, 2014]
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Case Study: GoogLeNet
[Szegedy et al., 2014]
Slide Credit: Fei Fei Li et al., Stanford
ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) winners
Slide Credit: Fei Fei Li et al., Stanford
Case Study:
ResNet
[He et al., 2015]
Slide Credit: Fei Fei Li et al., Stanford
Case Study:
ResNet
[He et al., 2015]
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Slide Credit: Fei Fei Li et al., Stanford
Object Detection
Object Detection using
Deep Learning
Bikash Santra
National Institutes of Health, USA
Object Detection
Slide Credit: Kaustav Kundu, University of Toronto
Object Detection
Slide Credit: Kaustav Kundu, University of Toronto
Object Detection
Slide Credit: Sihao Liang et al.
Object Segmentation Vs. Detection
Slide Credit: Kaustav Kundu, University of Toronto
Typical Object Detection Pipeline
Candidate Box Selection
Feature Extraction
Classification
Post processing
Slide Credit: Kaustav Kundu, University of Toronto
Typical Object Detection Pipeline
Slide Credit: Kaustav Kundu, University of Toronto
Evaluation Indicators for Object Detection
• Recall
• Precision
• mean Average Precision
(mAP)
• Intersection over Union (IoU)
Slide Credit: Sihao Liang et al.
Object Detection Competitions
Slide Credit: Sihao Liang et al.
Object Detection: Pre Deep Era
Slide Credit: Sihao Liang et al.
Deep Object Detection
1. Two Stage Detections
2. Unified Detections
Slide Credit: Sihao Liang et al.
Two Stage
Detections
Object Detection as Classification: Sliding Window
Slide Credit: Fei Fei Li et al. Stanford
Object Detection as Classification: Sliding Window
Slide Credit: Fei Fei Li et al. Stanford
Object Detection as Classification: Sliding Window
Slide Credit: Fei Fei Li et al. Stanford
Object Detection as Classification: Sliding Window
Slide Credit: Fei Fei Li et al. Stanford
Object Detection as Classification: Sliding Window
Slide Credit: Fei Fei Li et al. Stanford
Region Proposals / Selective Search
Slide Credit: Fei Fei Li et al. Stanford
RCNN : Region Proposal + CNN
1. Use selective search to come up with regional proposal
2. First object detection method using CNN
Slide Credit: Sihao Liang et al.
RCNN
Slide Credit: Sihao Liang et al.
RCNN
Slide Credit: Fei Fei Li et al. Stanford
RCNN
Slide Credit: Fei Fei Li et al. Stanford
RCNN
Slide Credit: Fei Fei Li et al. Stanford
RCNN
Slide Credit: Fei Fei Li et al. Stanford
RCNN
Slide Credit: Fei Fei Li et al. Stanford
RCNN
Slide Credit: Fei Fei Li et al. Stanford
Training RCNN
Slide Credit: Sihao Liang et al.
Training RCNN
Slide Credit: Sihao Liang et al.
Training RCNN
Slide Credit: Sihao Liang et al.
Training RCNN
Slide Credit: Sihao Liang et al.
Training RCNN
Slide Credit: Sihao Liang et al.
RCNN: Problems
Slide Credit: Fei Fei Li et al. Stanford
Thank you
