Match-Any-Events: Zero-Shot Motion-Robust Feature Matching Across Wide Baselines for Event Cameras

2026-04-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Match-Any-Events introduces the first event matching model capable of zero-shot, cross-dataset wide-baseline correspondence for event cameras, outperforming previous methods by 37.7%. Event cameras excel at instantaneous motion but struggle with wide-baseline matching due to appearance changes with motion and limited supervision. This new model features a motion-robust, computationally efficient attention backbone that learns multi-timescale features from event streams, enhanced by sparsity-aware event token selection. This design makes large-scale training on diverse wide-baseline supervision feasible. To address the data scarcity, the researchers developed a robust event motion synthesis framework to generate extensive event-matching datasets with augmented viewpoints, modalities, and motions. The model was trained on a combined large-scale dataset, achieving state-of-the-art results in semi-dense matching and camera pose estimation on both in-domain and unseen test data without fine-tuning.

Key takeaway

For research scientists developing event-based vision systems, Match-Any-Events demonstrates a significant leap in zero-shot wide-baseline matching. You should consider adopting its principles of separable spatial-temporal attention and sparsity-aware token selection to improve generalization and computational efficiency in your models. Furthermore, explore synthetic data generation frameworks like E-MegaDepth to overcome limitations in real-world wide-baseline supervision, enabling more robust and adaptable event camera applications.

Key insights

A new event matching model achieves zero-shot wide-baseline correspondence by combining efficient architecture and diverse synthetic data.

Principles

• Decouple spatial and temporal aggregation for efficiency.
• Adaptively prune redundant tokens to reduce computational cost.
• Synthesize diverse data for wide-baseline generalization.

Method

The method uses a Temporal Aggregation Transformer with separable spatial-temporal attention and Sparsity-aware Event Token Selection (SETS) on logarithmically windowed event voxels, followed by a Mutual Nearest Neighbors (MNN) matching stage.

In practice

• Utilize E-MegaDepth for synthetic wide-baseline training.
• Employ ECM dataset for real-world hetero-stereo evaluation.
• Integrate DINO pretrained weights for faster convergence.

Topics

• Event Cameras
• Zero-Shot Feature Matching
• Wide-Baseline Correspondence
• Spatiotemporal Transformers
• Sparsity-aware Token Selection

Code references

• spikelab-jhu/Match-Any-Events

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• NeuroLip: An Event-driven Spatiotemporal Learning Framework for Cross-Scene Lip-Motion-based Visual Speaker Recognition — Event cameras and specialized processing can enable robust lip-motion-based speaker recognition across diverse scenes.
• Metadata-Aware Multi-Prompt Reasoning for Zero-Shot Accident Understanding — Decomposing zero-shot video understanding into distinct stages improves reasoning reliability.
• SAM-Deep-EIoU: Selective Mask Propagation for Multi-Object Tracking — Selective mask propagation intelligently combines lightweight trackers with expensive VOS models to improve multi-object tracking efficiency and accuracy.
• Neural Events: Discrete Asynchronous Autoencoders for Event-Based Vision — The framework re-tokenizes high-resolution event streams into compressed, informative "neural events" using discrete learnable codes.
• FS-DVS: A Frequency-Selective Dynamic Visual Sensing Paradigm for Enhancing Information Completeness — FS-DVS uses a learnable pre-trigger spatial filter to mimic RGCs, enhancing event camera data completeness and noise resilience.
• EvTexture++: Event-Driven Texture Enhancement for Video Super-Resolution — EvTexture++ enhances VSR texture and temporal consistency by leveraging event signals beyond motion refinement.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.