Stratified Dense Matching for Stereoscopic Vision

 Jana Kostlivá (Kostková)
 Center for Machine Perception
Czech Technical University, Prague
http://cmp.felk.cvut.cz
 
  
Stereoscopic vision traditionally belongs to highly investigated topics in computer vision. The core of this filed is to establish correspondences between the input images. In our research we use stereo for 3D structure reconstruction. This application requires no false positive and mismatch errors. Consequently, the goal is to correctly match binocularly visible areas, detect half-occlusions, and identify unreliable regions (typically of low texture which are to be left unmatched). The Confidently Stable Matching algorithm is capable to solve this task. However, the quality of results is paid by lower matching density.

The matching density is directly related to the discriminability of matching features (e.g. contents of matching windows): the better the matching feature discriminability is the denser resulting disparity maps are. To  improve the discriminability, the matching windows have to cover the same parts of the scene structures. However, such features cannot be defined in the input images and their obtaining requires good quality matching hypothesis which have to be found in advance. Therefore, we propose a stratified dense matching approach.

We pose the problem in disparity space (which is a set of all tentative matches). The windows are defined to adapt to high-similarity structures in disparity space: disparity components (which represent matching hypothesis). The disparity within one disparity component (and thus also within the window) is allowed to vary in order to cope with non-planar surfaces and slanted objects. To this definition, two non-rectangular windows in the input images would typically correspond. The window definition in disparity space guarantees matching feature independence on projective distortions and invariance to input image views.

Stratified Dense matching Algorithm:

  1. Pre-matching: selects pre-matches (the high-similarity pairs)
  2. Disparity component tracing: traces out connected disparity components on pre-matches resulting from the first step. For each pre-match the corresponding disparity component, which directly determine the matching window, is identified uniquely.
  3. Similarity re-computation: computes a new similarity value for each pre-match based on the matching windows resulting from the second step.
  4. Final matching: finds the final matching using the new similarity values. For this step, the Confidently Stable Matching algorithm is used.


The Stratified Dense Matching approach improves matching feature discriminability, such that the density of results (false negative rate) is improved about 3x and the accuracy of results (mismatch rate) about 2.3x, comparing to the plain Confidently Stable Matching algorithm using fixed-size rectangular windows defined in the input images. The results on two very complex terrain scenes (the Larch Grove dataset and the Apple Tree dataset) are shown below. The results on standard datasets are shown here.
 
 

Larch Grove dataset
Apple Tree dataset

Confidently Stable Matching

Stratified Dense Matching


References:

Jana Kostkova
Last Modified: 08/04/2007