Richard Russell Graduate Student [email protected]
The Problems: Which computational image similarity metrics correspond most closely to human impressions of image similarity? How can image compression be optimized? What makes images look similar or dissimilar? How do humans perform comparisons of images?
Comparisons of different image similarity metrics L1 match Σ | xi-yi |
L2 match Σ (xi-yi)2
Vector quantization style image compression using different similarity metrics L1 Metric
What I like most about MIT/Cambridge: - ‘All under one roof’ - living in Boston - free espresso
What I like least about MIT/Cambridge: - expensive - winter - can’t buy booze on Sunday
Yuri Ostrovsky Graduate Student [email protected]
The Problems: What is the role of 3D in image processing? Does 3D scene information facilitate object search? How are different 3D scene attributes, such as illumination, represented by the visual system? Does object recognition influence 3D form perception?
Exploring the influence of 3D scene layout on visual search
Instant Good3D advantage
How fast do people find objects in Good3D vs. Bad3D?
Reaction Time (ms)
2400 2200 2000 1800 1600 1400 1200 1000 1
3 4 5 Trial Number for a Given Word
Amount of training Bad3D
Exploring the representation of 3D scene attributes such as illumination
What I like most about MIT/Cambridge: - The inexhaustible supply of nerdy rhetoric - Those charming long-time residents that scream at you for riding your bike on the sidewalk - Parades, parades, parades
What I like least about MIT/Cambridge: - Too much sunshine
Javid Q. Sadr Graduate Student [email protected]
The Problems: Can we develop a unified experimental paradigm for exploring different issues in object perception? What are the neural correlates of conscious object percepts? Can we develop novel, quantitative measures of object agnosias, development, and priming? How are object concepts learned, especially in the absence of supervision and image normalization?
The RISE Paradigm for Exploring Issues in Object Perception Idea of Object-Image Space and Random Image Structure Evolution....
... Phase RISE
... Bit-Flip RISE
A sample RISE sequence
What I like most about MIT/Cambridge: - the Rev. Sayan Mukherjee, Ph.D. - Top-Notch Local Weather Coverage - Dirty Deeds, Done Dirt Cheap - The Cambridgeport Saloon - Community-Based Policing
What I like least about MIT/Cambridge: - WHAT ! - SOMEONE SET UP US THE BOMB !
Jodi Davenport Graduate Student [email protected]
The Problems: How do we recognize faces? What aspects of a face do caricaturists capture to make super-recognizable portraits? What features are important for recognizing faces that are blurred or far away? How are objects represented in the brain?
Caricaturists have an intuitive sense of what features make a face recognizable as a unique individual. By creating a database of caricatures and real faces and developing computational schemes for extracting consistencies across the images of specific individuals, we hope to determine what aspects of a face define its recognizability.
Even when faces are far away or blurred and individual features (such as eyes, nose, lips) are difficult to discern, people are still able to recognize specific individuals. One goal is to define the heuristics used by human observers by determining the capabilities and limitations in face recognition with degraded images.
What I like most about MIT/Cambridge: • Students and Faculty with diverse interests. • Academic resources and talks. • Art loan program.
What I like least about MIT/Cambridge: • Tow trucks. • Lack of late night coffee-shops. • Slush.
Keith Thoresz Graduate Student [email protected]
The Problems: How can one extract automatically the invariant features of an arbitrary class of objects? Can a reliable object invariant be learned from a handful of examples? What kind of object model could be used to encode low resolution images to facilitate detection at a distance?
A qualitative invariant object representation. The ratio template is a compact object representation, robust to changes in illumination, noise and image degradation. Emphasizes lowfrequency (low-res) spectral components.
Detection results on faces. Note the range of spatial resolution and illumination differences.
What I like most about MIT/Cambridge: - New England sports (SCUBA, snowboarding) - Student-sponsored activities - Variety of lectures and talks institute-wide - Academic freedom and resources
What I like least about MIT/Cambridge: - No late-night food options - Classes
Antonio Torralba Postdoctoral Associate [email protected]
The Problems: What is the role of contextual information in object recognition? How can we efficiently represent context structure? How can we computationally model contextual influences on object recognition? What are the neural correlates of contextual processing?
Contextual influences in object recognition Object intrinsic features
Object recognition performances %
Object background features
Object in context %
A statistical framework for modeling contextual influences on object recognition Object priming:
Context driven focus of attention
Outdoor scene. Urban environment. Street. 50-100 meters.
Context driven scale selection
Car Person Traffic signs Furniture … categories
Location of Pedestrian area derived by context-based processing.
Likely pedestrian size derived via contextual cues
What I like most about MIT/Cambridge:
What I like least about MIT/Cambridge: - this guy