Repeatability experiment with real images

In this experiment we count the number of features extracted using the LB approach and compared them to the features extracted using SIFT. We use a real image and a pi/2 rotated version of this image. We count the number of points with a distance smaller than 2 pixels between the extracted features and the projection of the rotated points from the original image.

Original image SIFT

Rotated image SIFT

TOTAL POINTS = 650

Original image LB

Rotated image LB

TOTAL POINTS = 741

Matching using LB scale space on real images

In this experiment we use the same calibrated omnidirectional images but the scale space is computed using our Laplace-Beltrami operator. The first experiment shows the matching considering only rotation.

Matching using LB and Polar Descriptor

Matching using LB and SIFT Descriptor

The next matching experiment is performed between images with rotation and translation.

Matching using LB and Polar Descriptor

Matching using LB and SIFT Descriptor

Matching using real images and the polar descriptor

The first experiment consist of matching two omnidirectional images where one of them is obtained from the rotation of the first one around the z-axis by Pi/2.

Matching using Polar Descriptor

Matching using SIFT

The second experiment shows the matching between two different omnidirectional images, considering rotation and translation. A single octave is considered.

Matching using Polar Descriptor

Matching using SIFT

In the next experiment we use the four octaves.

Matching using Polar Descriptor

Matching using SIFT

We observe that the matching using LB through the scales cause the mismatching of the features, while that using the SIFT descriptor the matching is performed correctly. This indicates that the LB approach has problems with the matching throught the scales. More experiments have to be performed to identify the source of this behavior.

The next experiment matches the SIFT descriptors (128 vectors) using the QC criteria.

Matching SIFT using QC

Computing the support region on the sphere

We compute the support region on the sphere, which is required to compute the orientation and the descriptor.

Using the calibration, we map the extrema point detected to the sphere and define a vicinity proportional to the sin(sigma) where the point was detected. Then this vicinity is projected to the omnidirectional image and the orientation and magnitude gradients are computed. A weighted histogram is computed and its peak is selected as the orientation of the point.

Support region on the sphere

We observe that depending on the position on the sphere the support region in the image varies from circular to elliptical.

Support region for the descriptor

A similar process is followed to compute the descriptor corresponding to the detected feature. In this case the support region is divided in 36 bins. We verify the correctness of the orientation previously computed.

We can observe two examples of support regions over two omnidirectional images with the orientation previously computed

Good orientation

Image 1

Image 2

Bad orientation

Image 1

Image 2

This bad orientation computation can be caused because the orientation is computed with the gradients in the original image and not in the smoothed image where the feature was detected.

Matching using real images

In this experiment we perform the matching between two omnidirectional images acquired with an hypercatadioptric system. We compute the scale space using our approach, then this scale-space is passed to Vedaldi's software to compute the extrema points and the descriptors.

Extrema points features detected with LB



Normal SIFT points



Matching of LB features



Matching of SIFT features

The difference is explained since the scale-spaces used by the two approaches are different. The LB approach has smoothed images in the first two octaves but less smoothed in the las two. The opposite happens with Vedaldi's application, the first two octaves are less smtoothed and the last two are more smoothed than the LB ones.

Scale Space computed with LB

Normal SIFT Scale Space

Matching Experiments using Shape Context Locally

In this experiment we extract the edges of the images using Canny algorithm. For each extracted point we compute its corresponding support region based on the scale where the feature was detected. With all the edge pixels contained in this area we compute the histogram corresponding to that particular point. The approach used is taken from [1], with a log-polar grid of five ringsf and twelve sectors. The histogram-desciptor is a n x 60 matrix, where n is the number of edge points contained in the support region. The match process is performed using [2].

Edges with extrema points and descriptor example

Matching Results

We observe that some points have several matches which is not correct. This is possible caused by either the similarity matrix required by [2] or by the descriptor.

In the next experiment we change the descriptor. In this case the descriptor is only the 1x60 histogram counting the number of edge pixels in each cell. The results are similar to the previous case. We need to explore more the distance between the histograms we are using.

[1] G. Mori, S. Belongie, and J. Malik. "Shape Contexts Enable Efficient Retrieval of Similar Shapes".CVPR 2001
[2] Ofir Pele and Michael Werman. "The Quadratic-Chi Histogram Distance Family". ECCV 2010.

Matching Experiments using SIFT descriptors

In this experiments we use Vedaldi's software. We provide the scale space and the software computes the points and descriptors. Then using the same sotware we match the points.

Matching using LB scale space

Matching using SIFT

We observe that the SIFT implementation performs better that the one using the LB approach. This is because the scale space we computed has different scales from the ones that are used by the implementation. Even we change the number of octaves and scales, internally the initial scale is not changed. The descriptors use this scale and it might be the reason the are badly computed.