video: ego-centric and...

25
Video: Ego-centric and Summarization Presentation: Constance Clive Computer Science Department University of Pittsburgh

Upload: others

Post on 24-Jul-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Video: Ego-centric and Summarization

Presentation: Constance Clive

Computer Science Department

University of Pittsburgh

Page 2: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Nonchronological Video Synopsis and Indexing

Yael Pritch, Alex Rav-Acha, Shmuel Peleg

School of Computer Science and Engineering

The Hebrew University of Jerusalem

IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008

Page 3: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Motivation

• Effectively summarize activities from captured surveillance video

• Address queries on generated database objects

Page 4: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Approach

Page 5: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Results

• Online phase requires less than one hour to process an hour of vides (for typical surveillance video)

• Queries returned on the order of minutes depending on POI (Period of Interest)

Page 6: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Examples

• http://www.vision.huji.ac.il/video-synopsis

Page 7: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Detecting Activities of Daily Living in First-Person Camera Views

Hamed Pirsiavash, Deva Ramanan

Department of Computer Science, University of California, Irvine

Page 8: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

*slide courtesy of Piriavash and Ramanan

Page 9: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Motivation

• Tele-rehabilitation

• Life-logging for patients with memory loss

• represent complex spatial-temporal relationships between objects

• Provide a large dataset of fully annotated ADLs

Page 10: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Challenges long-scale temporal structure

time

Start boiling water

Do other things (while waiting)

Pour in cup Drink tea

Difficult for HMMs to capture long-term temporal dependencies

Wearable data: making tea

“Classic” data: boxing

*slide courtesy of Piriavash and Ramanan

Page 11: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Features

• Identify object:

• Aggregate features over time:

t = particular framei = a single objectp = pixel location and scaleT = set of frames to be analyzed

Page 12: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Temporal Pyramid

• Generate temporal pyramid

• Learn SVM classifiers on features for activity recognition:

= a histogram over a video clip

j = depth of the pyramid (level)

Page 13: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Temporal pyramidCoarse to fine correspondence matching with a multi-layer pyramid

Temporal pyramid

descriptor

Video clip

SVM

classifier

time

Inspired by “Spatial Pyramid” CVPR’06 and “Pyramid Match Kernels” ICCV’05

*slide courtesy of Piriavash and Ramanan

Page 14: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Active Object Models

• How to tell that an open fridge and a closed fridge are the same object?

• Train an additional object detector using the subset of “active” training images for a particular object

Page 15: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

“Passive” vs “active” objects

Passive Active

Page 16: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Dataset

• 20 people

• 30 minutes of footage a day

• 10 hours of footage per person

• 18 different identified ADLs

Page 17: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

ADL vs. Image-Net

Page 18: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Annotation

• 10 annotators, one annotation per 30 frames (1 second

• Action Label

• Object bounding box

• Object identity

• human-object interaction

• For co-occurring actions, the shorter interrupts the longer

Page 19: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Annotation

Page 20: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Functional Taxonomy

Page 21: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Experiment

• Leave-one-out cross-validation

• Average precision

• Class confusion matrices for classification error and taxonomy-derived loss

Page 22: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Training

• Off-the-shelf parts model for object detection

• 24 object categories

• 1200 training instances

• Inherent differences between training datasets:

Page 23: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Action Recognition results

Space-time interest points (STIP)Bag-of-objects model (O)Active-object model (AO)Idealized perfect object detectors (IO)Augmented Idealized object detectors (IA+IO)

Page 24: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while
Page 25: Video: Ego-centric and Summarizationpeople.cs.pitt.edu/~kovashka/cs3710_sp15/video_connie.pdfChallenges long-scale temporal structure time Start boiling water Do other things (while

Discussion

• Limitations?

• Future Work?