Q2. What are the future works in "Central catadioptric image processing with geodesic metric" ?
Moreover, it is worth noting that contrary to methods based on spherical harmonic analysis, the authors process points only in the image plane. It is also important to note that their approach can be applied to any type of single view point sensor, such as perspective camera, central catadioptric camera, but also fish-eye camera [ 7 ], since they all provide images equivalent to spherical images. Perspectives will then consist in applying these processing techniques to heterogeneous central sensor network in order to improve their efficiency and interest.
Q3. What are the intrinsic parameters of the sensor?
In order to take into account the distortions implied by the sensor during the omnidirectional image processing, the intrinsic parameters have to be considered and used for defining a new representation space.
Q4. How many corners are detected with the adapted Harris detector?
The geodesic Harris detector presents 43 common corners between images, i.e. a repeatability rate of 86%.20Feature matching between consecutive images of a sequence is a very important problem in computer vision for motion estimation for example.
Q5. What is the first step in the process of defining the convolution product?
the first step consists in defining the convolution product based on geodesic metric in order to process every pixel of the catadioptric image.
Q6. What is the spherical coordinates of the unit sphere?
The unit sphere S2 can be parameterized by spherical coordinates:∀x ∈ S2, x = (cos(φ) sin(θ), sin(φ) sin(θ), cos(θ)), (1)where φ ∈ [0, 2π[, θ ∈ [0, π].
Q7. Why are traditional image processing techniques no longer appropriate?
because of the distortions observed in such catadioptric images (fig 1), traditional image processing techniques are no longer appropriate and require to be adapted to the new sensor geometry.
Q8. How many corners are common between the two images?
In the classical Harris results, only 29 corners are common between the two images, which represents a repeatability equal to 80%.
Q9. How can the authors define a regular sampling at xs?
∑N i=−N ∑N j=−N IVNr (x)(i, j)H(i, j).(17)It is also possible to define a regular sampling at xs on the tangent plane π by using the distance dR2(x, y).
Q10. What is the way to apply the classical perspective metric?
It is also important to note that their approach can be applied to any type of single view point sensor, such as perspective camera, central catadioptric camera, but also fish-eye camera [7], since they all provide images equivalent to spherical images.
Q11. How many matchings are there in the classical neighborhood?
over the 21 images, the geodesic method provides 152 matchings in average while the classical methods gives 116 matchings.
Q12. What is the neighborhood of a pixel x in the image?
The authors then define the continuous neighborhood Vr(x) of pixel x in the image as follows:Vr(x) = {ys ∈ S2, dS2(xs, ys) 6 r |P(x) = xs}.
Q13. What is the neighborhood of the pixel x in the image?
In order to compare their results with conventional methods, the authors will set r such that the geodesic neighborhood is equivalent to the Euclidean neighborhood in the center of the catadioptric image where the distortions are negligible.
Q14. What is the rate of outliers in the neighborhood?
If the authors consider the same corners in both cases obtained by the classical Harris detector with the same thresholds, the use of the23classical ZNCC allows to match totally 65 points with 53 correct matchings, which represents a rate of outliers equal to 18.4%.