An efficient and low-latency attention model for event denoising.

Area of Science:

• Computer Vision
• Robotics
• Biomimetic Engineering

Background:

• Event-based cameras offer low-latency, high-temporal-resolution data.
• Background Activity (BA) noise corrupts event data, hindering performance.
• Existing denoising methods struggle with temporal dynamics and efficiency.

Purpose of the Study:

• Introduce StatFormer, an efficient event denoising method.
• Improve temporal feature modeling for event-based vision.
• Reduce noise in event camera data for enhanced downstream tasks.

Main Methods:

• Incorporated statistical information into event embedding for direct temporal modeling.
• Employed a two-stage transformer architecture with local attention.
• Focused on capturing fine-grained spatial-temporal dependencies and strengthening attention over sequences.

Main Results:

• StatFormer achieved a 13.41x speedup in inference.
• Demonstrated an average processing time of 1.38 μs per event.
• Delivered state-of-the-art denoising performance on public datasets.

Conclusions:

• StatFormer significantly enhances temporal feature modeling for event denoising.
• The method achieves superior performance with reduced model size and inference time.
• StatFormer offers a computationally efficient and effective solution for noisy event camera data.