Robust Event Based Vision Under Sensor Noise and Non Idealities

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2227 papers270 conceptsfrozen 2026-10-04

Event cameras, or Dynamic Vision Sensors (DVS), are often marketed as the "ultimate" vision sensors. By recording per-pixel intensity changes asynchronously rather than capturing global frames, they offer microsecond-level temporal resolution and a massive dynamic range. However, in practice, these sensors are notoriously "chatty." They don't just produce signals for motion; they produce a constant stream of noise that can overwhelm the very data researchers are trying to extract.

The research direction across these 168 papers focuses on characterizing, modeling, and mitigating this noise to make event-based vision viable for real-world applications like autonomous driving, high-speed tracking, and low-light photography.

The Anatomy of Event Noise

To understand why event cameras struggle, we must look at what triggers a "false" event. In a perfect world, an event is only generated when the log-intensity of light at a pixel changes beyond a threshold. In reality, as R. Wes Baldwin, Mohammed Almatrafi, and colleagues point out in their work on Event Probability Masks, circuitry leaks and thermal fluctuations create "background activity" (BA) noise.

This noise is not uniform. It scales aggressively with environmental factors:

  • Low Illumination: In the dark, the signal-to-noise ratio (SNR) of the sensor drops. Shih-Chii Liu and Tobi Delbruck demonstrate in the v2e toolbox that low light introduces parasitic events and increased latency, making it difficult to distinguish a moving object from the "thermal soup" of the sensor.
  • High Resolution: Counter-intuitively, more pixels can mean more problems. Daniel Gehrig and Davide Scaramuzza argue in Are High-Resolution Event Cameras Really Needed? that high-resolution sensors are hypersensitive to temporal noise, which can cause traditional model-based algorithms to break down.

The Flicker and Global Burst Problem

Beyond internal sensor noise, event cameras are hypersensitive to environmental "noise" that traditional cameras simply ignore. The most prominent is artificial light flicker. Because event cameras respond to any change in intensity, the 50/60Hz oscillation of AC-powered LEDs or fluorescent lights triggers a massive, periodic burst of events across the entire sensor.

Guillermo Gallego and Suman Ghosh highlight in their stereo vision survey that this flicker can obscure meaningful signal data. Similarly, Yunfan LU shows in the RGB-Event ISP benchmark that without specific flicker-mitigation strategies, multi-modal fusion models can actually perform worse than single-modality models because they end up "hallucinating" details based on the periodic noise.

Bridging the Reality Gap via Simulation

Because real-world event data is so messy, training deep learning models on it is difficult. This has led to a major research push in physically-accurate simulation.

The v2e framework by Yuhuang Hu and others is a cornerstone of this effort. Instead of just generating "clean" events from video frames, v2e models the non-idealities: intensity-dependent noise, motion blur, and the "dead time" of pixels. By training on "realistic" synthetic noise, researchers can create models that generalize to real hardware without being overwhelmed by the first hot pixel they encounter in the wild.

Cross-Modal Fusion as a Denoising Strategy

If the event stream is too noisy to stand alone, the solution is often to "guide" it using a secondary sensor—usually a standard RGB camera. This is the intuition behind Guided Event Filtering (GEF).

Zihao W. Wang proposed using the sharp spatial structure of a standard (but slow) intensity image to filter the high-speed (but noisy) event stream. By using the image as a prior, the system can distinguish between events that correspond to real edges and those that are just random fluctuations. This theme is echoed in CompEvent by Dong Li, which uses complex-valued fusion to handle low-light deblurring, and EV-Deblurformer by Kean Liu, which splits the task into temporal consistency and texture restoration.

From Motion Events to Illumination Maps

Perhaps the most radical shift in the field is moving away from treating events as "motion markers" and instead using them as "time-to-threshold" markers.

In RETINEV, Lei Sun and Yulun Zhang flip the script on low-light enhancement. Instead of waiting for an object to move, they use the camera's shutter itself to trigger "temporal-mapping" events. By measuring exactly how long it takes for each pixel to accumulate enough light to trigger its first event, they can reconstruct a dense illumination map. This method effectively turns the sensor's sensitivity into a strength, allowing for state-of-the-art restoration in environments so dark that traditional sensors see only black.

Papers in this version

Showing 60 of 227 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 198 papers
  2. Version 1 · 2026-10-04 · 227 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_event_sensor_noise_sensitivity_25ebefe9,
  title = {event sensor noise sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/event-sensor-noise-sensitivity/rdr_25ebefe9bd4749d2acc096e883ff5bb5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_event_based_vision_under_sensor_n_25ebefe9_v1,
  title = {Robust Event Based Vision Under Sensor Noise and Non Idealities},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-event-based-vision-under-sensor-noise-and-non-idealities/rdr_25ebefe9bd4749d2acc096e883ff5bb5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}