Research map / Image generation / Image restoration
LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds LED, a 3K-sequence real-world paired event denoising dataset whose ground truth comes from a dual-sampled event denoising framework (DED), and proposes DTSNN, a spiking network with dynamic LIF thresholds as a baseline.
Problem
Event cameras suffer from background activity (BA) noise, especially under low threshold and low illumination, but real paired event denoising data with trustworthy ground truth is scarce and small.
Approach
Two identical event cameras in a co-axial beamsplitter setup record the same scene; signal events are consistent across the two streams while BA noise is not, so their spatial intersection plus a spatiotemporal correlation filter yields clean GT. A fully spiking network whose LIF neurons use a learned dynamic threshold map then denoises events.
Key results
- DTSNN achieves 86.0 SR, 86.5 NR and 86.2 DA on LED, the best among compared methods (AEDNet 82.4 DA, EDnCNN 81.0 DA).
- The SNN version reaches 86.0 denoising accuracy versus 84.5 for its ANN counterpart, with an 18x lower energy ratio.
- DTSNN runs in 8.20 ms average, faster than AEDNet (708.1 ms) and EDnCNN (251.4 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DTSNN instead of AEDNet | event denoising · LED | DA: 82.4 -> 86.2 | Helped |
| use DTSNN instead of AEDNet | event denoising · DVS-CLEAN | SNR: 25.58 -> 29.26 | Helped |
| use DTSNN instead of AEDNet | event denoising · DVSNOISE20 | RPMD: 18.51 -> 16.13 | Helped |
| replace SNN neurons with ANN ReLU (ablation) | DTSNN · LED | Denoising accuracy: 86.0 -> 84.5 | Hurt |
| replace ANN ReLU neurons with spiking neurons (ablation) | DTSNN · LED | Energy ratio: 18× -> 1× | Helped |
| use fixed threshold 0.3 instead of dynamic threshold (ablation) | DTSNN · LED | Inference accuracy: 86.2 -> 85.5 | Hurt |
| use fixed threshold 0.5 instead of dynamic threshold (ablation) | DTSNN · LED | Inference accuracy: 86.2 -> 86.0 | Hurt |
| use fixed threshold 0.8 instead of dynamic threshold (ablation) | DTSNN · LED | Inference accuracy: 86.2 -> 85.7 | Hurt |
| remove spatial similarity from DED (ablation) | DED · LED | Denoising quality | Hurt |
| remove spatiotemporal correlation constraint from DED (ablation) | DED · LED | Denoising quality | Hurt |
Surprising findings
- The ANN version of the same architecture is less accurate than the SNN version while costing 18x more energy.
- Models trained on the synthetic DVSCLEAN dataset generalize poorly to real data, whereas models trained on the real LED data generalize well.
Limitations
- DED may fail when similar regions are dominated by noise events, though such cases are rare.
- DTSNN struggles to capture insignificant signals attached to structural peripheries, limiting signal retention.
- Lamp-induced signal events with noise-like appearance are retained as signal, which may look visually unappealing.
Experimental setup
Datasets: LED DVSCLEAN DVSNOISE20 DSEC E-MLB
Benchmarks: LED DVSCLEAN DVSNOISE20 DSEC E-MLB
Metrics: denoising accuracy (DA) signal retain (SR) noise removal (NR) SNR RPMD runtime
Baselines: Knoise DWF STDF TimeSurface EvFlow EDnCNN AEDNet
Builds on: LIF neuron model SpikingJelly E2VID
Closest papers by meaning
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.90)
- Mitigating Motion Blur in Neural Radiance Fields with Events and Frames (CVPR 2024 · similarity 0.87)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- Latency Correction for Event-guided Deblurring and Frame Interpolation (CVPR 2024 · similarity 0.87)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.86)
- Seeing Motion at Nighttime with an Event Camera (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.