Visual Scene Understanding (CS 598)

Post on 31-Dec-2015

28 views 1 download

Tags:

description

Visual Scene Understanding (CS 598). Derek Hoiem. Course Number: 46411 Instructor: Derek Hoiem Room:  Siebel Center 1109 Class Time:  Tuesday and Thursday 11:00am – 12:15pm Office Hours:  Tuesday and Thursday 12:15-1pm; by appointment Contact: dhoiem@uiuc.edu, Siebel 3312. Today. - PowerPoint PPT Presentation

Transcript of Visual Scene Understanding (CS 598)

Visual Scene Understanding (CS 598)

Derek Hoiem

Course Number: 46411Instructor: Derek HoiemRoom: Siebel Center 1109Class Time: Tuesday and Thursday 11:00am – 12:15pmOffice Hours: Tuesday and Thursday 12:15-1pm; by appointmentContact: dhoiem@uiuc.edu, Siebel 3312

Today

• Introductions

• Overview of logistics

• Overview of class material

Vision: What is it good for?

Biological (Humans)

1.2.3.4.5.6.7.8.9.10.

Technological (Computers)

1.2.3.4.5.6.7.8.9.10.

Note: Unfortunately, these got erased when my computer crashed

Course Logistics

Class Content Overview

• Tutorials and Perspectives

• Paper readingI) Spatial InferenceII) ObjectsIII) ActionsIV) Context and Integration

Visual Scene Understanding

Visual scene understanding is the ability to infer

general principles and current situations from imagery in a way that helps achieve goals.

Visual Scene Understanding

Visual scene understanding is the ability to infer

general principles and current situations from imagery in a way that helps achieve goals.

Visual Scene Understanding

Visual scene understanding is the ability to infer

general principles and current situations from imagery in a way that helps achieve goals.

Visual Scene Understanding

Visual scene understanding is the ability to infer

general principles and current situations from imagery in a way that helps achieve goals.

I. Spatial Inference

Getting Around

Getting Around

Getting Around

Spatial Inference: applications

Household RobotsAutomated Vehicles

Graphics ApplicationsPredict object size/position

Spatial Inference: open questions

• How do we represent space?– Surface orientations, depth maps, voxels?

• How do we infer it from available sensory data (image, stereo, motion, laser range finder)?

II. Objects

Finding Things and Observing Them

Image classification: Are there any dogs?Photo credit: iansand – flickr.com

Finding Things and Observing Them

Object Localization: Where are the dog(s)?

Finding Things and Observing Them

Verification: Is this a dog?

Finding Things and Observing Them

Description: Furry, small, nice, side view

Finding Things and Observing Them

Identification: My friend Sally?

Recognizing Stuff

SKY

SAND

WATER

Object Recognition: applications

Photo SearchSecurity

Robots

Object Recognition: open questions

• How many examples does it take to learn one category well?

• How many examples does it take to learn 100 categories well?

• How do these answers depend on the level of supervision?

• Can recognition be solved with simple methods and massive amounts of data?

• How can we quickly recognize an object?

• How can we scale up to deal with thousands of categories?

III. Actions

Taking Action

[Saxena et al. 2008]

Recognizing Actions

KTH Dataset

Figure from Laptev et al. 2008

Recognizing Actions

Figure from Laptev et al. 2008

Reading Emotions

Photo credit: Comstok

Actions: applications

SecurityVideo Search

Actions: open questions

• How are actions defined?

• Does it make sense to categorize them?– If not, how do we recognize them?

• What are good visual representations for inferring actions?

• How can we recognize activities?

IV. Context and Integration

[Hoiem et al. 2008]

Context and Integration

[Hoiem et al. 2008]

• Objects + scene categories better detection

• Movement + objects action/activity recognition

• Space + objects navigation

Context and Integration: applications

Everything that vision is good for

Context and Integration: open questions

• Should context be explicit (e.g., “cars drive on the road”) or implicit (feature-based)?

• How do we model and learn the interactions between different processes and scene characteristics?

• How do we deal with the growing complexity as more and more pieces are put together?

General Problems in Computer Vision

• Better understanding of limitations and their sources– Need new experimental paradigms

• Improve generalization– Aim to generalize across datasets, categories, and

tasks– Work on knowledge sharing and transfer

• Vision as a way of learning about the world– Integration into AI– Systems that acquire knowledge over time

Successes of Computer Vision• Point matching (e.g. 2d3)

– Tracking– Structure from motion– Stitching

• Product inspection• Multiview 3d reconstruction• Face recognition and modeling• Object recognition on pre-2000 datasets• Interactive segmentation (ongoing)

To Do

• Register on bulletin board

• Post comments on Thursdays reading (due tomorrow)

• Look over schedule and decide which days to present (due next Tues)

• Start thinking about projects– Let me know if you want a specific pairing (due Tues)

Questions?

Goals

• Make you a better researcher (esp. in vision)– More knowledge– Better critical thinking skills– Improved communication skills– Improved research skills

Grades

• Participation: 25%– Posting– Class discussion

• Presentation: 25%

• Projects: 50%– Proposal, progress report, final paper, and oral

Policies

• Attendance required (see syllabus)

• Give credit where due

• No formal prerequisites

• Everything needs to be on time

Reading

• Read well

• Post comments to bulletin board at least 24 hours before class

Presentations• Presenter

– Everyone does two– Good quality coverage of topic (40 min)– See syllabus for guidelines– Sign up by next Tuesday (at latest)– TBAs are your choice (decide at least 4 weeks in advance)

• Demonstrator– If all days are taken, pair up– One person’s job will be to demonstrate some aspect of the algorithm

(e.g., where it succeeds and fails) by running it on many examples– May require implementation

• Note taker

Projects• Timeline

– Proposal: Feb 12 (3 ½ weeks!)– Progress report: Mar 19– Presentation: paper May 5, oral later

• Progress report• Presentation

– Paper– Oral

• In pairs– Can choose partner or be randomly paired

• Suggestions on web

• Potentially will lead to publication (e.g. NIPS)

To Do

• Register on bulletin board

• Post comments on Thursdays reading (due tomorrow)

• Look over schedule and decide which days to present (due next Tues)

• Start thinking about projects– Let me know if you want a specific pairing (due Tues)

Questions?