Skip to content

Topic 8: Optic flow

  1. Read and discuss the paper by Lucas and Kanade. Try to capture the essentials in less than half a page of the report.

  2. Relate the approach of the paper to the methods presented in the textbook by Trucco and Verri [2] and in the methods presented in the lecture notes (slides).

  3. Read and discuss the paper by Yan Niu et al [3]. Try to capture the essentials in less than half a page of the report.

  4. What are the restrictions of the proposed technique in [1] in terms of changes in:

    • motion (deformation),
    • object shape,
    • object intensity.

  5. The size of the local image (window) corresponds to the summation term Σx in many equations. Is there a minimal and a maximal size? Explain your answer.

  6. Implement the technique by Lucas and Kanade [1] using Matlab and DIPimage.

  7. Use your implementation to study the behaviour of the presented iterative method for large displacements (for which the method still works). How many iterations are required and what is a suitable stopping criteria?

    Hint

    Is there a limit on the precision that can be reached? If so, what is the limiting factor?

  8. Use your own implementation to find the translation vector between two shifted images. What is the maximum displacement for which the method works? Can you relate this to the size of the image?

  9. Investigate the influence of selecting reliable feature points using the confidence measure proposed in [3] on the flow computation.

  10. Implement the method for computing Optic Flow (see lecture notes on Registration, Motion, optic flow and eq. (8.25) of Trucco and Verri [2] using Matlab and DIPimage. Please note that equation for computing optic flow in the lecture notes is identical to the one presented by [2].

  11. Compute the flow field of a synthetic image sequence called “rotating_points”. This image is of size 128x128x60, indicating 60 time frames of size 128x128 pixels. Please present the flow field of frame 30 in a vector plot and specify your choice of filter parameters (sigma of Gaussian derivatives in x,y, and t (time) and the size of the tensor smoothing).

References

[1] Bruce D. Lucas, Takeo Kanade, An Iterative Image Registration Technique with Application to Stereo Vision, Proceedings DARPA Image Understanding Workshop, pp121-130, April 1981.

[2] E. Trucco and A. Verri, Introductory Techniques for 3-D Computer Vision, Prentice-Hall, 1998.

[3] Yan Niu, Zhiwen Xu and Xiangjiu Che, Dynamically Removing False Features in Pyramidal Lucas-Kanade Registration, IEEE Trans. On Image Processing, Vol 23, no 8, August 2014

Last update: 2023-04-17