[CV][Detector][Contour] gPb

1. Introduction

- gPB is state of the art contour detector [2] proposed by Michael Maire et al. at CalTech.

- Bryan Catanzaro et al (Berkeley) re-implemented this detector on CUDA [1].

They contributed a parallel version with rotated integral image for histogram calculation and used Lanczos algorithms with Cullum-Willoughby test for eigensolver to speed-up about 100 times (4 mins down to 1.8 secs).

They claimed in [1] that their implementation gives the same result as in [2] - 0.7 on Berkeley dataset (Figure 6 and Figure 7 is results) and tested on several GPU kinds (Table 2).

2. Dataset

They tested on Benchmark Dataset in Countour detection [Berkeley Segmentation Dataset]:

3. Source code:

- gPb: Software and source code can be downloaded from Michael Maire's page

- Damascene: parallel gPb on CUDA by ParLab at Beckeley)

4. My conclusion:

- The result of gPb seems good and Catanzaro's implementation takes the advantages of speed-up ---> should try

References:

[1] Bryan Catanzaro, Narayanan Sundaram, Bor-Yiing Su, Yunsup Lee, Mark Murphy, Kurt Keutzer. "Efficient, High-Quality Image Contour Detection". In 2009 IEEE 12th International Conference on Computer Vision (September 2009)

[2] M. Maire, P. Arbelaez, C. Fowlkes, and J. Malik. "Using contours to detect and localize juncitons in natural images". CVPR, pages 1-8, June 2008.

[3] Contour detecton survey slide by Michael Maire.

[4] Michael Maire's thesis on Machael Maire's page