Insights into Batch Selection for Event-Camera Motion Estimation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Event cameras provide high temporal resolution data streams, ideal for motion estimation.
• Unlike frame-based cameras, event data is asynchronous, posing challenges in data aggregation.
• Defining optimal event batching strategies is crucial for computational tasks.

Purpose of the Study:

• Investigate optimal event batch selection for neural network-based rotational motion estimation.
• Determine the impact of different event batching methods on motion estimation performance.
• Evaluate the hypothesis that a minimum event count is necessary for motion estimation.

Main Methods:

• Utilized neural networks for rotational motion estimation using event camera data.
• Experimented with various event batching strategies (e.g., fixed-time windows, fixed-count batches).
• Analyzed the influence of batch selection on training and inference performance.

Main Results:

• Event batch selection significantly impacts motion estimation results.
• Training requires diverse event batches, irrespective of the selection method.
• Fixed-time windows are effective for inference and comparable to complex methods.
• The hypothesis of a minimal event count for motion estimation was invalidated for neural networks.

Conclusions:

• Neural network performance in rotational motion estimation is sensitive to event batching.
• Diverse batch selection during training is essential for robust performance.
• Fixed-time window batching offers a practical and effective approach for inference.