1: \begin{abstract}

2: Event cameras are novel sensors that naturally respond to the scene dynamics and offer advantages 

3: to estimate motion.

4: Seeking the accuracy gains that image-based deep-learning methods have achieved, 

5: optical flow estimation methods for event cameras have rushed to combine those image-based methods 

6: with ideas from top performing event-based frameworks.

7: However, it requires several adaptations (data conversion, loss design, etc.) 

8: as events and images have very different properties,

9: and a supervisory signal that matches the capabilities of event data.

10: We develop a principled method to extend the Contrast Maximization framework to estimate optical flow from events alone.

11: We do so by investigating key elements: 

12: how to design the objective function to prevent overfitting, 

13: how to warp events to deal better with occlusions, %

14: and how to improve convergence without discretizing events.

15: With these key elements, our method ranks first on the MVSEC indoor benchmark,

16: and is competitive on the MVSEC outdoor and the DSEC benchmarks.

17: Moreover, our method allows us to expose the issues of the ground truth flow in current datasets and benchmarks,

18: and produces remarkable results when it is transferred to unsupervised learning architectures.

19: \end{abstract}

20: