The method significantly improves the state of the art for accuracy on three common dataset: Stanford Background (8 classes), SIFT Flow (33 classes) and Barcelona (170 classes).
The method is also roughly 100 times faster than the best competitor: roughly 0.5 second per frame on a 4-core laptop. An FPGA implementation of the system (based on the NeuFlow architecture) runs the bulk of the algorithm in 50 ms at 1280x240 resolution, using about 10 Watts of power.
It is available here.
My paper on ultrametric watersheds is amongst the selected 5 top-cited articles from the Journal of Mathematical Imaging and Vision available free until the end of September.
More information in the Springer NewsLetter
The paper is avalable here. This HAL version will be forever freely available.
Another prize for Camille Couprie! She has just been awarded an accessit for the prestigious Gilles Khan prize.
Congratulations, Camille!
Shape based filtering
Non-Local Dual Constrained Total Variation Denoising (NL-DCTV)
Saliency maps
Scene parsing with hierarchical convolutional nets
The Polygonal Path Image (PPI)