Research

My research interests include classical computer vision problems including tracking, detection, segmentation, and feature design, as well as problems specific to microscopic imagery including tree reconstruction, denoising, registration, and alignment. I am also interested in developing active learning methods to speed up and improve classification on large data sets.

SLIC Superpixels SLIC Superpixels
Recently, superpixels have become increasingly popular. They are a useful primitive for feature extraction, capture redundancy, and reduce complexity of subsequent processing tasks. SLIC superpixels are extremely fast, simple, and achieve better segmentation quality than state-of-the-art methods.
[paper | code]
 
CRF with global constraints High Order CRF Segmentation
Top performing segmentation algorithms rely on CRF models that enforce spatial and global consistency through additional latent variables. Our experiments on the PASCAL and the MSRC datasets show that similar performance can be acheived with a much simpler design that ignores such constraints. [details | paper]
 
MCMC tracking of migrating neurons Tracking Migrating Neurons
We developed a Markov Chain Monte Carlo (MCMC) framework for detecting and tracking hundreds of neurons migrating in time-lapse 2-photon microscopy. Our approach exploits a correlated motion and appearance model to improve performance (neurons elongate in the direction of travel).
[details | paper | poster | demo | video]
 
Detecting Neurons and Mitochondria Detection using Ray Features
Rays are a powerful group of image features designed to consider image characteristics at distant contour points. Used with an Adaboost classifier, they can reliably detect neuron nuclei and mitochondria. Rays can characterize deformable or irregular shapes with consistent responses and can be efficiently precomputed.
[further details | paper | poster]
 
Mitochondria Segmentation Mitochondria EM Segmentation
Despite success on natural images, conventional segmentation algorithms fail on EM data. We developed a fully automatic graph partitioning approach to segment mitochondria. It uses global shape cues and learns to model the appearance of membranes.
[further details | paper | demo]
 
3D Mitochondria Segmentation 3D EM Segmentation
EM image stacks can contain tens of millions of voxels. Segmenting 3D mitochondria from such large volumes requires an efficient segmentation method, supervoxel partitioning, and global shape cues to recognize the appearance of a mitochondria.
[further details | video | video | report]
 
DIADEM Tree Reconstructions DIADEM Dendrite Reconstruction
Our DIADEM team placed 4th in the DIADEM Challenge out of 125 submissions, and was awarded a $10,000 prize. The goal of the competition was to develop automatic methods to reconstruct dendritic arbors from various types of microscopic data.
[further details | video | video]
 
VIVA visualization VIVA Visualization & Analysis
Visualizing large biomedical data is a difficult challenge. We are currently developing VIVA - an open source Volumetric Visualization and Analysis tool, which can render, annotate, and analyze extremely large and high dimensional data.
[further details | video]
kevin smith

Kevin Smith
Assistant Professor
School of Computer Science and Communication

Email: ksmith (at) kth.se

Address:
Science for Life Laboratory
KTH Royal Institute of Technology
School of Computer Science and Communication
Box 1031
SE-17121 Solna
Sweden

Phone:+46 8 790 64 37

 
Valid HTML 4.01 Strict