Will you be doing dual tvl1?

[hollowgalaxy]

Or can you give some pointers on how to alter your code to match dual tvl1?

[rfezzani]

Hi @hollowgalaxy, the pyimof.solvers.tvl1 is already the dual TV-L1algorithm (see the doc) ;-)

[hollowgalaxy]

Awesome. Thanks. Really helpful repo.

[rfezzani]

You welcome ;-) Thank you @hollowgalaxy for your interest in pyimof

[hollowgalaxy]

I finally had a chance to compare your implementation with the one in cv2 using this code...

import matplotlib.pyplot as plt
import pyimof
import cv2
import numpy as np

print(cv2.__version__) # 4.1.0
I0, I1 = pyimof.data.hydrangea()
dof = cv2.optflow.createOptFlow_DualTVL1()
flow = dof.calc(I0=np.expand_dims(I0, axis=-1), I1=np.expand_dims(I1, axis=-1), flow=None)
u = flow[...,0]
v = flow[...,1]
# u, v = pyimof.solvers.tvl1(I0, I1)

pyimof.display.plot(u, v)
plt.show()

There is a slight difference.
Here is cv2's solution

Here is your solution (showing here for easy comparison)

Seems that the cv2 implimenation is more coarse. My thought would have been that this signifies its better. Or does smoother pixel transition indicate better results?

[rfezzani]

hi @hollowgalaxy, obtaining sharp optical flow at object border is a great challenge. In fact these regions are usually the occlusion areas, (ie areas where the optical flow problem is undefined).

The answer to your question basically depends on your application: if you want to use the optical flow for segmentation for example, sharp results are better. But if you want to estimate fluid motion (in PIV for example), there is no edges to fit, so smooth optical flow is OK.

Now if you look at the pyimof.data.urban2 sequence for example, you will see that sharp result does not always mean good result: in fact opencv output is sharp but not good at all:

compared to pyimov

and the ground truth optical flow:

[hollowgalaxy]

Wow. Awesome answer. I appreciate the detail you went through. I was wondering why that noise was appearing in the cv2 implementation. Looks like it was just that, noise.