Abstract
Online optimisation revolves around new data being introduced into a problem while it is still being solved; think of deep learning as more training samples become available. We adapt the idea to dynamic inverse problems such as video processing with optical flow. We introduce a corresponding predictive online primal-dual proximal splitting method. The video frames now exactly correspond to the algorithm iterations. A user-prescribed predictor describes the evolution of the primal variable. To prove convergence we need a predictor for the dual variable based on (proximal) gradient flow. This affects the model that the method asymptotically minimises. We show that for inverse problems the effect is, essentially, to construct a new dynamic regulariser based on infimal convolution of the static regularisers with the temporal coupling. We finish by demonstrating excellent real-time performance of our method in computational image stabilisation and convergence in terms of regularisation theory.
Similar content being viewed by others
Notes
The total variation term in (1.3) in principle requires \(x \in {\mathrm {BV}}(\varOmega )\), the space of functions of bounded variation on \(\varOmega \). This is not a Hilbert space, but merely a Banach space, where our overall setup (1.1) does not to apply. However, due to the weak(-\(*\)) lower semicontinuity of convex functionals, any minimiser of (1.3) necessarily lies in \(L^2(\varOmega ) \cap {\mathrm {BV}}(\varOmega )\), so we are justified in working in the Hilbert space \(X=L^2(\varOmega )\), and seeing \({\mathrm {BV}}(\varOmega )\) as a constraint imposed by the total variation term.
Any coercive, convex, proper, lower semicontinuous function \(E: X \rightarrow {\overline{\mathbb {R}}}\) has a minimiser \({\hat{x}}\). By the Fermat principle \(0 \in \partial E({\hat{x}})\). Thus, \({\hat{x}}\in \partial E^*(0)\), which says exactly that \(E^* \ge E^*(0)\).
The double arrow signifies that the map is set-valued.
Then (5.4) gives \(\varLambda _\mathcal {V}=1\). Constant true displacements are allowed by Lemma 5.1, but constant measurements not. If \(\Vert x-{\bar{x}}\Vert _{L^2(\varOmega + B(0, \Vert u\Vert ))}^2 \le C \Vert x-{\bar{x}}\Vert _{L^2(\varOmega )}^2\) then Lemma 5.1 and Theorem 5.1 extend to \(\varLambda > C\varLambda _\mathcal {V}\). In practise, to compute \(x \circ v\), we extrapolate x outside \(\varOmega \) such that Neumann boundary conditions are satisfied.
To obtain the linearised optical flow model, we start with \(b_{k+1}(\xi )=b_{k}(v_k(\xi ))\) holding for all \(\xi \in \varOmega \) and a sufficiently smooth image \(b_{k}\). By Taylor expansion \(b_{k}(v_k(\xi )) \approx b_{k}(\xi ) + \langle \nabla b_{k}(\xi ),v_k(\xi )-\xi \rangle \). Thus \(0 = b_{k+1}(\xi )-b_{k}(v_k(\xi )) \approx b_{k+1}(\xi )-b_{k}(\xi ) + \langle \nabla b_{k}(\xi ),\xi -v_k(\xi )\rangle \).
References
Aragón Artacho, F.J., Geoffroy, M.H.: Characterization of metric regularity of subdifferentials. J. Convex Anal. 15(2), 365–380 (2008)
Bastianello, N., Simonetto, A., Carli, R.: Prediction-correction splittings for time-varying optimization with intermittent observations. IEEE Control Syst. Lett. 4(2), 373–378 (2020). https://doi.org/10.1109/LCSYS.2019.2930491
Bauschke, H.H., Combettes, P.L.: Convex Analysis and Monotone Operator Theory in Hilbert Spaces, 2 edn. CMS Books in Mathematics. Springer (2017). https://doi.org/10.1007/978-3-319-48311-5
Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J. Imag. Sci. 2(1), 183–202 (2009). https://doi.org/10.1137/080716542
Becker, F., Petra, S., Schnörr, C.: Optical flow. In: O. Scherzer (ed.) Handbook of Mathematical Methods in Imaging, pp. 1945–2004. Springer (2015). https://doi.org/10.1007/978-1-4939-0790-8_38
Belmega, E.V., Mertikopoulos, P., Negrel, R., Sanguinetti, L.: Online convex optimization and no-regret learning: Algorithms, guarantees and applications (2018)
Bezanson, J., Edelman, A., Karpinski, S., Shah, V.B.: Julia: A fresh approach to numerical computing. SIAM Rev. 59(1), 65–98 (2017)
Biegler, L., Ghattas, O., Heinkenschloss, M., Keyes, D., Waanders, B.: Real-Time PDE-Constrained Optimization. SIAM, Computational Science and Engineering (2007)
Bousquet, O., Bottou, L.: The tradeoffs of large scale learning. Adv. Neural Inf. Process. Syst. 20, 161–168 (2008)
Chambolle, A.: An algorithm for total variation minimization and applications. J. Math. Imaging. Vis. 20(1), 89–97 (2004). https://doi.org/10.1023/B:JMIV.0000011325.36760.1e
Chambolle, A., Pock, T.: A first-order primal-dual algorithm for convex problems with applications to imaging. J. Math. Imaging Vis. 40, 120–145 (2011). https://doi.org/10.1007/s10851-010-0251-1
Chaudhury, K., Mehrotra, R.: A trajectory-based computational model for optical flow estimation. IEEE Trans. Robot. Autom. 11(5), 733–741 (1995). https://doi.org/10.1109/70.466611
Chen, K., Lorenz, D.A.: Image sequence interpolation based on optical flow, segmentation, and optimal control. IEEE Trans. Image Process. (2012). https://doi.org/10.1109/TIP.2011.2179305
Clason, C., Valkonen, T.: Introduction to nonsmooth analysis and optimization (2020). https://tuomov.iki.fi/m/nonsmoothbook_part.pdf. Work in progress
Engl, H., Hanke, M., Neubauer, A.: Regularization of Inverse Problems. Springer, Mathematics and Its Applications (2000)
Franzen, R.: Kodak lossless true color image suite. PhotoCD PCD0992. Lossless, true color images released by the Eastman Kodak Company (1999). http://r0k.us/graphics/kodak/
Grötschel, M., Krumke, S., Rambau, J.: Online Optimization of Large Scale Systems. Springer, New York (2013)
Hall, E., Willett, R.: Dynamical models and tracking regret in online convex programming. In: S. Dasgupta, D. McAllester (eds.) Proceedings of the 30th International Conference on Machine Learning, Proceedings of Machine Learning Research, vol. 28, pp. 579–587. PMLR, Atlanta, Georgia, USA (2013)
Hazan, E.: Introduction to online convex optimization. Found. Trends Optim. 2(3–4), 157–325 (2016). https://doi.org/10.1561/2400000013
Horn, B.K., Schunck, B.G.: Determining optical flow. In: Proc. SPIE, vol. 0281, pp. 319–331. SPIE (1981). https://doi.org/10.1117/12.965761
Iglesias, J.A., Kirisits, C.: Convective regularization for optical flow, pp. 184–201. De Gruyter (2016). https://doi.org/10.1515/9783110430394
Jauhiainen, J., Kuusela, P., Seppänen, A., Valkonen, T.: Relaxed Gauss-Newton methods with applications to electrical impedance tomography. SIAM J. Imaging Sci. 13, 1415–1445 (2020). https://doi.org/10.1137/20M1321711
Nagel, H.H.: Extending the ‘oriented smoothness constraint’ into the temporal domain and the estimation of derivatives of optical flow. In: Faugeras, O. (ed.) Computer Vision–ECCV 90, pp. 139–148. Springer, Berlin (1990)
Nagel, H.H., et al.: Constraints for the estimation of displacement vector fields from image sequences. In: Proceedings of the Eighth International Joint Conference on Artificial Intelligence (II), vol. 2, pp. 945–951. IJCAI (1983)
Orabona, F.: A modern introduction to online learning (2020)
Salgado, A., Sánchez, J.: Temporal constraints in large optical flow estimation. In: R. Moreno Díaz, F. Pichler, A. Quesada Arencibia (eds.) Computer Aided Systems Theory–EUROCAST 2007, pp. 709–716. Springer, Berlin (2007)
Simonetto, A.: Time-varying convex optimization via time-varying averaged operators (2017)
Tico, M.: Digital image stabilization. In: A.A. Zaher (ed.) Recent Advances in Signal Processing, chap. 1. IntechOpen, Rijeka (2009). https://doi.org/10.5772/7458
Valkonen, T.: Transport equation and image interpolation with SBD velocity fields. Journal de mathématiques pures et appliquées 95, 459–494 (2011). https://doi.org/10.1016/j.matpur.2010.10.010
Valkonen, T.: Testing and non-linear preconditioning of the proximal point method. Applied Mathematics and Optimization (2018). https://doi.org/10.1007/s00245-018-9541-6
Valkonen, T.: First-order primal-dual methods for nonsmooth nonconvex optimisation (2019). https://tuomov.iki.fi/m/firstorder.pdf. Submitted
Valkonen, T.: Julia codes for “predictive online optimisation with applications to optical flow”. Software on Zenodo (2020). https://doi.org/10.5281/zenodo.3659180
Valkonen, T.: Preconditioned proximal point methods and notions of partial subregularity. Journal of Convex Analysis (2020). (in press)
Volz, S., Bruhn, A., Valgaerts, L., Zimmer, H.: Modeling temporal coherence for optical flow. In: 2011 International Conference on Computer Vision, pp. 1116–1123. IEEE (2011). https://doi.org/10.1109/ICCV.2011.6126359
Weickert, J., Schnörr, C.: Variational optic flow computation with a spatio-temporal smoothness constraint. J. Math. Imaging Vis. 14(3), 245–255 (2001). https://doi.org/10.1023/A:1011286029287
Zhang, Y., Ravier, R.J., Tarokh, V., Zavlanos, M.M.: Distributed online convex optimization with improved dynamic regret (2019)
Zhou, J., Hubel, P., Tico, M., Schulze, A.N., Toft, R.: Image registration methods for still image stabilization. US Patent 9,384,552 (2016)
Zinkevich, M.: Online convex programming and generalized infinitesimal gradient ascent. In: Proceedings of the 20th International Conference on Machine Learning (ICML-03), pp. 928–936. AAAI (2003)
Acknowledgements
This research has been supported by Escuela Politécnica Nacional internal Grant PIJ-18-03 and Academy of Finland Grants 314701 and 320022.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Local Strong Convexity
Local Strong Convexity
We establish local strong convexity of the indicator function of the ball. This has been shown in [1] to be equivalent to the strong metric subregularity of the subdifferential. For related characterisations, see also [33] and regarding total variation [22, appendix].
Lemma A.1
With \(F: X \rightarrow \mathbb {R}\), \(F=\delta _{{{\,\mathrm{cl}\,}}B(0, \alpha )}\) on a Hilbert space X, suppose \(x \in \partial B(0, \alpha )\) and \(0 \ne x^* \in \partial F(x)\). Then,
for
Proof
Observe that \(x^*=\lambda x\) for \(\lambda :=\Vert x^*\Vert /\alpha \). If \(x' \not \in {{\,\mathrm{cl}\,}}B(0, \alpha )\), there is nothing to prove. So take \(x' \in {{\,\mathrm{cl}\,}}B(0, \alpha )\). Then, we need \( 0 \ge \lambda \langle x,x'-x\rangle + \frac{\gamma }{2}\Vert x'-x\Vert ^2. \) Since \(\Vert x\Vert =\alpha \), this says
Suppose \(\gamma \le \lambda \), which is the first case of \(U_x\). Then (A.1) is seen to hold by application of Young’s inequality on the inner product term, followed by \(\Vert x'\Vert \le \alpha \).
If on the other hand, \(\gamma > \lambda \), which is the second case of \(U_x\), we take \(x' \in {{\,\mathrm{cl}\,}}B(x, \alpha ) \cap {{\,\mathrm{cl}\,}}B(0,\alpha )\). This implies \( \langle x',x\rangle \ge \tfrac{1}{2}\Vert x'\Vert ^2. \) Since \(\lambda -\gamma <0\), this and \(\Vert x'\Vert \le \alpha \) prove (A.1). \(\square \)
Rights and permissions
About this article
Cite this article
Valkonen, T. Predictive Online Optimisation with Applications to Optical Flow. J Math Imaging Vis 63, 329–355 (2021). https://doi.org/10.1007/s10851-020-01000-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10851-020-01000-4