Research map / Architectures and efficiency / RNNs and normalization
EventPillars: Pillar-based Efficient Representations for Event Data
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 EventPillars, a pillar-based dense event representation combining temporal event range, activated event polarity, and normalized event density, achieving state-of-the-art recognition and detection accuracy with lower computation and storage.
Problem
Dense image-like representations of sparse asynchronous event-camera data either lose temporal, polarity, or count information or are computationally expensive, hindering edge deployment of event-based vision.
Approach
Adapts pillar-based point-cloud encoding to events: per-pixel pillars store newest and earliest timestamps (Temporal Event Range), a continuous activated polarity balance (AEP), and a max-normalized event density (NED) used as a spatial attention prior, all concatenated into a 4-channel tensor.
Key results
- EventPillars reaches 97.12% accuracy on N-Cars and 85.3% on N-Caltech101, 52.3% on N-ImageNet, surpassing EST and others.
- On object detection EventPillars achieves 53.1% mAP on Gen1 and 44.4% mAP on 1 Mpx with augmentation, versus 50.4% and 40.6% for ERGO-12.
- Representation time is 0.62 ms per 100 ms N-Cars sample, about 9.2x lower than EST and 3.9x lower than DiST, with 4.5x lower input storage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace EST with EventPillars | ResNet-34 recognition pipeline · N-Caltech101 | accuracy: 81.7 -> 85.3 | Helped |
| replace EST with EventPillars | ResNet-34 recognition pipeline · N-Cars | accuracy: 92.5 -> 97.1 | Helped |
| replace EST with EventPillars | ResNet-34 recognition pipeline · N-ImageNet | accuracy: 48.9 -> 52.3 | Helped |
| replace ERGO-12 with EventPillars, with augmentation | Swin V2 with YOLOv6 head · Gen1 | mAP: 50.4 -> 53.1 | Helped |
| replace ERGO-12 with EventPillars, with augmentation | Swin V2 with YOLOv6 head · 1 Mpx | mAP: 40.6 -> 44.4 | Helped |
| add earliest timestamp (EET) to newest timestamp (NET) (ablation) | N-Cars temporal representation · N-Cars | accuracy: 89.23 -> 92.86 | Helped |
| use EET only instead of NET only (ablation) | N-Cars temporal representation · N-Cars | accuracy: 89.23 -> 88.95 | Hurt |
| use tanh activation with α=0.4 for AEP (ablation) | N-Cars polarity representation · N-Cars | accuracy: 85.57 | No clear effect |
| replace AEP with Event Histogram (ablation) | N-Cars polarity representation · N-Cars | accuracy: 85.57 -> 86.54 | Helped |
| normalize event counts by the maximum pillar count to form NED (ablation) | N-Cars density representation · N-Cars | accuracy: 81.32 -> 86.06 | Helped |
| add AEP to NET (ablation) | N-Cars EventPillars components · N-Cars | accuracy: 89.23 -> 91.33 | Helped |
| add AEP to NET + EET (ablation) | N-Cars EventPillars components · N-Cars | accuracy: 92.86 -> 95.92 | Helped |
Surprising findings
- The Event Histogram baseline with implicit polarity outperforms the explicit AEP alone (86.54% vs. 85.57%), which the authors attribute to richer count information.
Limitations
- Only a single temporal window is explored; alternative window sizes and temporal bin partitioning are not studied.
- Ablation studies are run only on N-Cars to limit training cost.
Experimental setup
Datasets: N-Cars N-Caltech101 N-ImageNet Gen1 1 Mpx MS-COCO
Benchmarks: N-Cars N-Caltech101 N-ImageNet Gen1 1 Mpx
Metrics: accuracy mAP representation time (ms) input size
Baselines: HOTS Event Histogram Event Image HATS Voxel Grid DiST EST ERGO-12 TORE MDES
Builds on: PointPillars TinyPillarNet ERGO-12 EST DiST
Closest papers by meaning
- ESEG: Event-Based Segmentation Boosted by Explicit Edge-Semantic Guidance (AAAI 2025 · similarity 0.88)
- EventZoom: A Progressive Approach to Event-Based Data Augmentation for Enhanced Neuromorphic Vision (AAAI 2025 · similarity 0.87)
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 2025 · similarity 0.86)
- Event-assisted Low-Light Video Object Segmentation (CVPR 2024 · similarity 0.86)
- Event-based Structure-from-Orbit (CVPR 2024 · similarity 0.86)
- EventGPT: Event Stream Understanding with Multimodal Large Language Models (CVPR 2025 · similarity 0.85)
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.