Learning Efficient Meshflow and Optical Flow From Event Cameras.

Area of Science:

• Computer Vision
• Robotics
• Event-based Sensing

Background:

• Event cameras offer high temporal resolution and low latency for motion estimation.
• Existing methods for event-based flow estimation lack meshflow-specific datasets and struggle with varying event data density.
• Accurate motion field prediction is crucial for real-time applications in robotics and autonomous systems.

Purpose of the Study:

• To address the limitations in current event-based meshflow estimation techniques.
• To introduce a novel dataset and a lightweight network for efficient and accurate meshflow prediction.
• To investigate and improve the robustness of event-based methods across varying data densities.

Main Methods:

• Generation of the High-Resolution Event Meshflow (HREM) dataset with 1280x720 resolution, dynamic objects, complex motion, and optical/meshflow labels.
• Proposal of the Efficient Event-based MeshFlow (EEMFlow) network, a lightweight encoder-decoder model for swift meshflow estimation.
• Development of a Confidence-induced Detail Completion (CDC) module for dense event optical flow and an Adaptive Density Module (ADM) to enhance generalization.

Main Results:

• EEMFlow demonstrates exceptional performance and is 30x faster than state-of-the-art methods.
• The HREM dataset provides a comprehensive benchmark for meshflow estimation.
• The ADM module significantly improves EEMFlow and EEMFlow+ performance by 8% and 10% respectively, showcasing enhanced robustness across varying event densities.

Conclusions:

• The proposed EEMFlow network offers a significant advancement in efficient and accurate event-based meshflow estimation.
• The HREM dataset and ADM module contribute to addressing key challenges in event-based motion analysis.
• This work paves the way for more robust and generalized event-based motion perception systems.