zhu_ev-flownet_2018: EV-FlowNet: self-supervised optical flow estimation for event-based cameras
EV-FlowNet: self-supervised optical flow estimation for event-based cameras
Event data is well-suited for Optical Flow Estimation
By directly measuring the precise time at which the pixel changes, the event stream encodes fine-grained motion information. For example, optical flow can be estimated from a local window around each event in a linear fashion.
The authors propose a novel representation of event data, as an image with channels representing the number of events, and the latest timestamp at each polarity at each pixel. This preserves spatial relationships between events, while maintaining the most recent temporal information at each pixel, and providing a fixed number of channels for any event stream.
They also introduced a Self-supervised Learning method, using a photometric loss as a supervisory signal, without supervision from ground-truth flow.
The event stream is represented as a 4 channel image:
- Number of positive polarity events
- Number of negative polarity events
- Timestamp of most recent positive event
- Timestamp of most negative event
Park, Paul K. J., Baek Hwan Cho, Jin Man Park, Kyoobin Lee, Ha Young Kim, Hyo Ah Kang, Hyun Goo Lee, et al. n.d. “Performance Improvement of Deep Learning Based Gesture Recognition Using Spatiotemporal Demosaicing Technique.” In 2016 IEEE International Conference on Image Processing (ICIP), 1624–28.