Reproduce “DA4Event: towards bridging the Sim-to-Real Gap for Event Cameras using Domain Adaptation” on OpenResearch

About this paper

Mirco Planamente, Chiara Plizzari, Marco Cannici, Marco Ciccone, Francesco Strada, Andrea Bottino, Matteo Matteucci, Barbara Caputo

Researchers introduced Unsupervised Domain Adaptation (UDA) and a novel Multi-View architecture (MV-DA4E) to address the sim-to-real gap for event cameras, allowing deep learning models trained on simulated event data to perform effectively on real-world event streams. The approach yielded consistent performance gains of up to 11% and showed superior generalization compared to methods adjusting simulator parameters.

Key insights

• Applying Unsupervised Domain Adaptation (UDA) at the feature level effectively mitigates the sim-to-real gap for event cameras, proving more robust than solely adjusting simulator parameters.
• The Multi-View DA4E (MV-DA4E) architecture consistently enhances performance across various event representations by preserving ImageNet pre-trained weights and fusing information from multiple 3-channel views.
• Event representations that maintain temporal resolution tend to be more inherently robust to domain shifts compared to those that condense event information into single frames.

Summary by alphaXiv.