Research map / Video / Video segmentation and tracking
EventZoom: A Progressive Approach to Event-Based Data Augmentation for Enhanced Neuromorphic Vision
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.
Proposes EventZoom, a progressive scaling-and-shifting event data augmentation that embeds zoomed event samples into others to preserve spatial integrity and temporal continuity, evaluated in supervised, semi-supervised and unsupervised SNN settings on DVS-CIFAR10, N-Caltech101 and UCF101-DVS.
Problem
Existing event-data augmentation methods (CutMix-style cutting and replacement, mixup) break spatial integrity and temporal continuity of event streams, hurting classification performance on limited neuromorphic datasets.
Approach
Embed randomly scaled and shifted event sequences into an original sample, with scale and position interpolated linearly over time steps, so no cropping occurs and labels are soft-weighted per time step by the occupied area.
Key results
- EventZoom reaches 85.90% on DVS-CIFAR10 with Spike-VGG11 at 48x48 resolution and 85.75% on N-Caltech101 with TET training, outperforming listed augmentation baselines.
- On N-Caltech101 with Pre-Act ResNet-18, EventZoom improves accuracy from 67.93 (no augmentation) to 78.39, a +10.46 gain.
- In semi-supervised DVS-CIFAR10 with 40 labels per class, EventZoom achieves 71.60 (Pre-Act ResNet-18) vs 57.00 without augmentation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EventZoom instead of EventMix | Pre-Act ResNet-18 · DVS-CIFAR10 | accuracy: +0.80 | Helped |
| Use EventZoom instead of EventMix | Pre-Act ResNet-18 · N-Caltech101 | accuracy: +7.94 | Helped |
| Use EventZoom instead of EventMix | Pre-Act ResNet-18 · UCF101-DVS | accuracy: +4.12 | Helped |
| Add EventZoom augmentation instead of no augmentation | Pre-Act ResNet-18 · DVS-CIFAR10 (40 labels) | accuracy: +20.60 | Helped |
| Add EventZoom augmentation instead of no augmentation | Pre-Act ResNet-18 · DVS-CIFAR10 (100 labels) | accuracy: +6.60 | Helped |
| Add EventZoom augmentation instead of no augmentation | Pre-Act ResNet-18 · DVS-CIFAR10 (250 labels) | accuracy: +3.50 | Helped |
| Add EventZoom instead of EventMix | Shallow-spiking-VGG11 · N-Caltech101 pretrain → N-Caltech101 fine-tune | accuracy: +1.72 | Helped |
| Add EventZoom instead of EventMix | Shallow-spiking-VGG11 · N-Caltech101 pretrain → DVS-CIFAR10 fine-tune | accuracy: +5.10 | Helped |
| Increase the λ range from 0.0–0.0 to 0.2–0.6 (ablation) | EventZoom on Shallow-Spiking-VGG11 · N-Caltech101 | accuracy: -0.23 | Hurt |
| Increase the λ range from 0.0–0.0 to 0.3–0.7 (ablation) | EventZoom on Shallow-Spiking-VGG11 · N-Caltech101 | accuracy: +0.80 | Helped |
| Increase the λ range from 0.0–0.0 to 0.3–1.0 (ablation) | EventZoom on Shallow-Spiking-VGG11 · N-Caltech101 | accuracy: +3.45 | Helped |
| Increase the λ range from 0.0–0.0 to 0.4–1.3 (ablation) | EventZoom on Shallow-Spiking-VGG11 · N-Caltech101 | accuracy: +5.98 | Helped |
Surprising findings
- Random scaling and shifting were reported to disrupt temporal continuity and cropping to compromise spatial integrity, yet the paper's ablation is limited to one dataset and model.
- Per the paper, MixUp can disrupt illumination information in event data and perform worse than no augmentation in some settings.
Limitations
- Only tested on classification tasks; downstream tasks not evaluated due to lack of suitable event datasets.
- Labels are averaged across time steps for computational convenience rather than using per-step soft labels.
- Mixing more samples (mixnum) may add noise and complexity.
Experimental setup
Datasets: DVS-CIFAR10 N-Caltech101 UCF101-DVS DvsGesture Bullying10K
Benchmarks: DVS-CIFAR10 N-Caltech101 UCF101-DVS DvsGesture Bullying10K
Metrics: accuracy training cost time (ms)
Baselines: MixUp CutMix EventMix EventDrop NDA EventRPG ShapeAug EventAugment
Builds on: CutMix Mixup EventMix EventDrop NDA
Closest papers by meaning
- Leveraging Asynchronous Spiking Neural Networks for Ultra Efficient Event-Based Visual Processing (AAAI 2025 · similarity 0.87)
- EventGPT: Event Stream Understanding with Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
- PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.87)
- EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.87)
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 2025 · similarity 0.87)
- V2V: Scaling Event-Based Vision through Efficient Video-to-Voxel Simulation (NeurIPS 2025 · similarity 0.87)
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.