An opposition to Window- Scanning Approaches in Computer Vision Presented by Tomasz Malisiewicz March 6, 2006 Advanced Perception @ The Robotics Institute.

An opposition to Window-Scanning Approaches in

Computer Vision

Presented by Tomasz Malisiewicz

March 6, 2006Advanced Perception @ The Robotics Institute

2 Problems

• Does scanning windows across an image work?

• What types of objects does it work for?

What are window-scanning approaches missing?

*Following Slides Borrowed From Derek Hoiem’s “Putting Context Into Vision” Presentation

Contextaka Top-Down Processing

Quick Question: What is this?

What is context?

• Any data or meta-data not directly produced by the presence of an object– Nearby image data

Context

What is context?

• Any data or meta-data not directly produced by the presence of an object– Nearby image data– Scene information

Context

Context

What is context?

• Any data or meta-data not directly produced by the presence of an object– Nearby image data– Scene information– Presence, locations of other objects

Tree

Clues for Function

• What is this?

Clues for Function

• What is this?

• Now can you tell?

Low-Res Scenes

• What is this?

Low-Res Scenes

• What is this?

• Now can you tell?

More Low-Res

• What are these blobs?

More Low-Res

• The same pixels! (a car)

Why is context useful?

• Objects defined at least partially by function

– Trees grow in ground – Birds can fly (usually)– Door knobs help open doors

Why is context useful?

• Objects defined at least partially by function– Context gives clues about function

• Not rooted into the ground not tree• Object in sky {cloud, bird, UFO, plane,

superman} • Door knobs always on doors

Why is context useful?

• Objects defined at least partially by function– Context gives clues about function

• Objects like some scenes better than others

• Toilets like bathrooms• Fish like water

Why is context useful?

• Objects defined at least partially by function– Context gives clues about function

• Objects like some scenes better than others

• Many objects are used together and, thus, often appear together

• Kettle and stove• Keyboard and monitor

The other* problem

• What types of objects does it work for?

*Assuming we can just directly avoid the first problem

• “Our goal is to develop a system that detects and recognizes many kinds of objects in photographs and video including everyday office objects, text captions in video, and various structures in biomedical imagery.” – Schneiderman and Kanade from Object Detection Using the Statistics of Parts

How many different classifiers must one construct?A different classifier for each object?

A different classifier for each pose of an object?How many poses do we need per object?

“However, such approaches seem unlikely to scale up to the detection of hundreds or thousands of different object classes because each

classifier is trained and run independently.” – Torralba and Murphy and Freeman

from Sharing features: efficient boosting procedures

for multiclass object detection

Too many windows

• Now imagine scanning a window and applying 100K independent classifiers at each window

Conclusion

• Without context, we can’t find all things we want to find. We need context to help constrain the search for objects.

• With independent classifiers per object (and per pose), we can’t detect a large number of objects. Should cow detectors and a horse detectors be built independently? Think along the lines of a horse and a cow are types of animals that often occur in similar contexts.

• Remember that complex and deformable objects would require many poses if are to adhere to the window-based classifier paradigm.

Thank you.

*Pascal 2006 Visual Challenge Image