Segmentation for Robot navigation
Members : Sang Min Oh, Frank Dellaert, James M. Rehg.
Primary Contact : Sang Min Oh (sangmin (with at) cc.gatech (dot edu).
1. Running Fast-Segmentation (by Felzenswalb and Huttenlocher) on Lagr images (2006/ 2/ 18)
-- For Indoor and Outdoor images. Robot vision images are fast-segmented.
[Link]
2. Running Auto Popup (by Derek Hoeim, Efros, Herbert) on Lagr images (2006/ 2/ 25)
-- For Indoor and Outdoor images. Auto-popup predicts each super-pixel's geometric context.
[Link]
3. Swendsen Wang Cut Ground-Truth Comparison (2006/ 3/ 7)
-- Swendsen Wang Cut is implemented and compared to Ground-truth to validate its correct behavior.
[Link]
4. Swendsen Wang Cut on Color Gradient Images (2006/ 3/ 13)
-- Color Gradient images are (1) fast-segmented, (2) MRFs are built on the over-segmentations, then
(3) SWC algorithm is run on the MRFs and (4) we see the results.
[Link]
[NOT MAINTAINED ANYMORE, E-MAIL ME FOR THE RESULTS]
5. Swendsen Wang Cut on Outdoor LAGR robot images (2006/ 3/ 16)
-- Lagr Outdoor images are (1) fast-segmented, (2) MRFs are built on the over-segmentations, then
(3) SWC algorithm is run on the MRFs and (4) we see the results.
+ All the measurement models for MRFs are learned autonomously by the robot.
[Link]
6. Swendsen Wang Cut on Outdoor LAGR robot Video (2006/ 3/ 27)
-- (1) Color models of Traversable / Non-Traversable areas are being learned online., then
(2) Each input image is segmented using SWC technique (an MCMC). (3) Segment the images into sky,
traversalbe/non-traversable areas. (4) The posterior probability is displayed using a color coding
blud for sky, red for non-traversable and green for traversable areas. (5) Later, the classification
results are fed into the planner, and we check whether the segmentation improves planning capabilities.