Find relative scale in monocular Visual Odometry without PnP - python

I am implementing a standard VO algorithm, with a few changes, i.e. extract features, match the feature points, find the essential matrix and decompose to get the pose. After initialization however, instead of using 3D-2D motion estimation (PNP) for subsequent frames, I'm using the same 2D-2D motion estimation (using essential matrix). I find that 2D-2D estimation seems a lot more accurate than 3D-2D.
To find the relative scale of the second pose with respect to the first, I can find out the common points (that were triangulated for both frame pairs). According the Visual Odometry Tutorial, Scaramuzza, one can find the relative scale by finding the ratio of relative distances between common point pairs.
If f13D and f23D are the triangulated 3D points from subsequent framepairs, I choose point pairs in random and compute the distances, here is a rough code snippet for the same.
indices = np.random.choice(np.arange(0,len(f23D)), size=(5 * len(f23D),2),replace=True)
indices = indices[indices[...,0]!=indices[...,1]]
num = np.linalg.norm(f13D[indices[...,0]] - f13D[indices[...,1]], axis=1).reshape((len(indices),1))
den = np.linalg.norm(f23D[indices[...,0]] - f23D[indices[...,1]], axis=1).reshape((len(indices),1))
return np.median(num/den).
I have also tried replacing the last line with a linear ransac estimator. However, since scale triangulation is not perfect, these values are extremely noisy and thus the scale estimate also varies significantly, on using different numpy seeds.
Is this the right way to implement relative scale in monocular VO as described in the article? If not, what is the best way to do it (I do not wish to use PNP since rotation seems to be less accurate)

Related

Constraining RBF interpolation of 3D surface to keep curvature

I've been tasked to develop an algorithm that, given a set of sparse points representing measurements of an existing surface, would allow us to compute the z coordinate of any point on the surface. The challenge is to find a suitable interpolation method that can recreate the 3D surface given only a few points and extrapolate values also outside of the range containing the initial measurements (a notorious problem for many interpolation methods).
After trying to fit many analytic curves to the points I've decided to use RBF interpolation as I thought this will better reproduce the surface given that the points should all lie on it (I'm assuming the measurements have a negligible error).
The first results are quite impressive considering the few points that I'm using.
Interpolation results
In the picture that I'm showing the blue points are the ones used for the RBF interpolation which produces the shape represented in gray scale. The red points are instead additional measurements of the same shape that I'm trying to reproduce with my interpolation algorithm.
Unfortunately there are some outliers, especially when I'm trying to extrapolate points outside of the area where the initial measurements were taken (you can see this in the upper right and lower center insets in the picture). This is to be expected, especially in RBF methods, as I'm trying to extract information from an area that initially does not have any.
Apparently the RBF interpolation is trying to flatten out the surface while I would just need to continue with the curvature of the shape. Of course the method does not know anything about that given how it is defined. However this causes a large discrepancy from the measurements that I'm trying to fit.
That's why I'm asking if there is any way to constrain the interpolation method to keep the curvature or use a different radial basis function that doesn't smooth out so quickly only on the border of the interpolation range. I've tried different combination of the epsilon parameters and distance functions without luck. This is what I'm using right now:
from scipy import interpolate
import numpy as np
spline = interpolate.Rbf(df.X.values, df.Y.values, df.Z.values,
function='thin_plate')
X,Y = np.meshgrid(np.linspace(xmin.round(), xmax.round(), precision),
np.linspace(ymin.round(), ymax.round(), precision))
Z = spline(X, Y)
I was also thinking of creating some additional dummy points outside of the interpolation range to constrain the model even more, but that would be quite complicated.
I'm also attaching an animation to give a better idea of the surface.
Animation
Just wanted to post my solution in case someone has the same problem. The issue was indeed with scipy implementation of the RBF interpolation. I tried instead to adopt a more flexible library, https://rbf.readthedocs.io/en/latest/index.html#.
The results are pretty cool! Using the following options
from rbf.interpolate import RBFInterpolant
spline = RBFInterpolant(X_obs, U_obs, phi='phs5', order=1, sigma=0.0, eps=1.)
I was able to get the right shape even at the edge.
Surface interpolation
I've played around with the different phi functions and here is the boxplot of the spread between the interpolated surface and the points that I'm testing the interpolation against (the red points in the picture).
Boxplot
With phs5 I get the best result with an average spread of about 0.5 mm on the upper surface and 0.8 on the lower surface. Before I was getting a similar average but with many outliers > 15 mm. Definitely a success :)

Python get transformation matrix from two sets of points

I have to images, one simulation, one real data, with bright spots.
Simulation:
Reality:
I can detect the spots just fine and get the coordinates. Now I need to compute transformation matrix (scale, rotation, translation, maybe shear) between the two coordinate systems. If needed, I can pick some (5-10) corresponding points by hand to give to the algorithm
I tried a lot of approaches already, including:
2 implementations of ICP:
https://engineering.purdue.edu/kak/distICP/ICP-2.0.html#ICP
https://github.com/KojiKobayashi/iterative_closest_point_2d
Implementing affine transformations:
https://math.stackexchange.com/questions/222113/given-3-points-of-a-rigid-body-in-space-how-do-i-find-the-corresponding-orienta/222170#222170
Implementations of affine transformations:
Determining a homogeneous affine transformation matrix from six points in 3D using Python
how to perform coordinates affine transformation using python? part 2
Most of them simply fail somehow like this:
The red points are the spots from the simulation transformed into the reality - coordinate system.
The best approach so far is this one how to perform coordinates affine transformation using python? part 2 yielding this:
As you see, the scaling and translating mostly works, but the image still needs to be rotated / mirrored.
Any ideas on how to get a working algorithm? If neccessary, I can provide my current non-working implementations, but they are basically as linked.
I found the error.
I used plt.imshow to display both the simulated and real image and from there, pick the reference points from which to calculate the transformation.
Turns out, due to the usual array-to-image-index-flipping-voodoo (or a bad missunderstanding of the transformation on my side), I need to switch the x and y indices of the reference points from the simulated image.
With this, everything works fine using this how to perform coordinates affine transformation using python? part 2

Python : detect dominant wavelength of a signal

I have this signal, for which I want to calculate the dominant wavelength, which would be the distance between the pronounced minima where the oscillations occure:
Which tool in scipy should I look into for this mission?
It depends where you get the data from.
If you only have the (x,y) points of the graph, you can either hack it by taking all the x corresponding to the minimal y (be careful of floating-point equalities, though), or use the Fourier transform, identify the main wave (the biggest amplitude) and deduce its wavelength. For the latter, you would use the Fast Fourier Transform from scipy: https://docs.scipy.org/doc/scipy-0.18.1/reference/tutorial/fftpack.html#fast-fourier-transforms
If you have the functional description of the function, either sample it like you do to construct the graph and apply the above, or take its derivative to find the minima mathematically (best method). You could also use scipy to find the minima numerically (https://docs.scipy.org/doc/scipy-0.18.1/reference/generated/scipy.optimize.minimize.html), but you have to manually specify intervals that contain only one local minimum.

How to improve HOG detector with linear SVM performance for car detection?

So, I want to detect cars from a driver recorder recorded video. I've read a lot and do research quite a lot but still not quite getting it. I do think of using a HOG descriptor with linear SVM. But in what way it can still be improver to make it easier to be implement and more robust since this will be kind of a research for me?
I am thinkin of combining another technique/algorithm with the HOG but still kind of lost. I am quite new in this.
Any help is greatly appreciated. I am also open to other better ideas.
HOG (histogram of oriented gradients) is merely a certain type of feature vector that can be computed from your data. You compute the gradient vector at each pixel in your image and then you divide up the possible angles into a discrete number of bins. Within a given image sub-region, you add the total magnitude of the gradient pointing in a given direction as the entry for the relevant angular bin containing that direction.
This leaves you with a vector that has a length equal to the number of bins you've chosen for dividing up the range of angles and acts as an unnormalized histogram.
If you want to compute other image features for the same sub-region, such as the sum of the pixels, some measurement of sharp angles or lines, aspects of the color distribution, or so forth, you can compute as many or as few as you would like, arrange them into a long vector as well, and simply concatenate that feature vector with the HOG vector.
You may also want to repeat the computation of the HOG vector for several different scale levels to help capture some scale variability, concatenating each scale-specific HOG vector onto the overall feature vector. There are other feature concepts like SIFT and others, which are created to automatically account for scale invariance.
You may need to do some normalization or scaling, which you can read about in any standard SVM guide. The standard LIBSVM guide is a great place to start.
You will have to be careful to organize your feature vector correctly since you will likely have a very large number of components to the feature vector, and you have to ensure they are always calculated and placed into the same ordering and undergo exactly the same scaling or normalization treatments.

Find the most significant corner of a skeleton and segment the skeleton at that corner

I have images of ore seams which I have first skeletonised (medial axis multiplied by the distance transform), then extracted corners (see the green dots). It looks like this:
The problem is to find a turning point and then segment the seam by separating the seam at the turning point. Not all skeletons have turning points, some are quite linear, and the turning points can be in any orientation. But the above image shows a seam which does have a defined turning point. Other examples of turning points look like (using ASCII): "- /- _". "X" turning points don't really exist.
I've tried a number of methods including downsampling the image, curve fitting, k-means clustering, corner detection at various thresholds and window sizes, and I haven't figured it out yet. (I'm new to to using scikit)
The technique must be able to give me some value which I can use heuristically determine whether there is a turning point or not.
What I'd like to do is to do some sort of 2 line ("piecewise"?) regression and find an intersection or some sort of rotated polynomial regression, then determine if a turning point exists, and if it does exist, the best coordinate that represents the turning point. Here is my work in progress: https://gist.github.com/anonymous/40eda19e50dec671126a
From there, I learned that a watershed segmentation with appropriate label coordinates should be able to segment the skeleton.
I found this resource: Fit a curve for data made up of two distinct regimes
But I wasn't able to figure out to apply it my current situation. More importantly there's no way for me to guess a-priori what the initial coefficients are for the fitting function since the skeletons can be in any orientation.

Categories

Resources