contrib/gel/gevd/gevd_contour.h File Reference

tracing connected contours and junctions More...

#include <vcl_vector.h>
#include <vbl/vbl_array_2d.h>
#include <vtol/vtol_vertex_2d_sptr.h>
#include <vtol/vtol_edge_2d_sptr.h>
#include <gevd/gevd_bufferxy.h>

Go to the source code of this file.

Classes
class	gevd_contour

---

Detailed Description

tracing connected contours and junctions

Operator to implement the tracing of connected contours and junctions. All contours are assumed chains or cycles with zero width, and so can be traced using 4/8-connected pixels. Junctions will be detected from end points touching multiple other end points, or touching a stronger contour with a significant jump in filter response. Extensively use the image grid to insure planarity of the network of edges and vertices, when projected onto the image plane. The recipe is:

1. Trace 4/8-connected pixels to enumerate disjoint chains and cycles. Prune contours that are either short or have no pixel stronger than a high hysteresis threshold.

2. Find junctions from end points touching internal pixels of some stronger chain/cycle. Break a stronger contour at a junction, only if there is a detectable jump in filter response. Next, merge end points touching other end points or junctions. Finally, create dummy end point for isolated cycles.

3. Insert subpixel accuracy into edges/vertices. Because of truncation errors, the mapping from edgel locations to integral grid locations may no longer be preserved, after this step.

4. Optionally reduce noisy zig-zags along the contours, and evenly space the contour points. The zig-zags are never more than 0.5 pixel, and happen when sub-pixel locations are noisy and so out-of-sync with the 4/8-connected tracing.

5. Optionally insert virtual border at the image boundary, to form closed region beyond the image boundary.

6. Insert depth/z values into edges/vertices, through interpolation of the range image, for example. For an intensity image, set this depth to a constant value, since the 3D edges/vertices all lie in the image plane.

Input: connected edge elements, with response strength, and subpixel location, describing isolated contours disjoint only at junction pixels. min_strength and min_length are used to prune weak or short edges. min_jump is used to prune weak junctions.

Output: planar network of linked edges and vertices.

Complexity: O(|edgels|) time and space. O(nlogn) time for quicksort if n=|chains| < 1000, to make sure that junctions are found from longer/stronger chains first.

Author:

John Canny (1986) SM Thesis Chris Connolly (1987) Use Fu-Tsao thinning Van-Duc Nguyen (1989) Eliminate short & weak contours Arron Heller (1992) Translate from CLOS to C++ Van-Duc Nguyen (1995) Trace/search breadth-first instead of thinning Joe Mundy (1997) Added continuous edgel orientation output Van-Duc Nguyen (1998) Merge from end points of dangling chains only Joe Mundy (1999) Modified ::InsertBorder to use ROI bounds Peter Vanroose (2003) removed z coord arg of Translate()

Definition in file gevd_contour.h.