Efficient Multi-Timescale Event Representations for Feed-Forward Object Detection

Fredrik Lundell; Per-Erik Forssen; Mårten Wadenbäck; Astrid Lundmark

cs.CVSep 4, 2026

Efficient Multi-Timescale Event Representations for Feed-Forward Object Detection

Authors: Fredrik Lundell, Per-Erik Forssen, Mårten Wadenbäck, Astrid Lundmark

Organizations: Department of Electrical Engineering, Linköping University, Sweden · Saab AB, Linköping, Sweden

Abstract

Autonomous systems require robust low-latency perception under rapidly changing scene dynamics and challenging illumination. In event cameras object detection commonly relies on recurrent architectures to accumulate sparse temporal information over time. This work investigates how temporal information can be encoded directly within the event representation. We propose a confidence-normalized continuous multi-timescale representation based on logarithmic B-spline temporal encoding together with a geometry-aware local confidence mechanism that exploits the spatial structure of event generation. Using a fixed feed-forward EventCenterNet detector, we show that the proposed representations consistently outperform the compact CSTR representation on PEDRo and Gen1 datasets. We further introduce a recursive exponential-polynomial approximation that enables efficient event-by-event updates while largely preserving detection performance. These results demonstrate that carefully designed event representations can capture a substantial portion of the temporal information learned through recurrent temporal modeling, providing a promising foundation for efficient feed-forward, event-driven, and future neuromorphic object detection.

Figures & tables

Panel of Figure 1 — Figure 1: Temporal structure of the proposed logarithmic B-spline representation. (a) Four Second-order logarithmic B-spline functions for $\tau_{0}=1.2$ and $\tau_{\max}=35$ . (b) Correlation matrix of logarithmic B-spline (left) and exponential decay (right) temporal functions, showing that the B-spline functions produce a less redundant multi-timescale encoding.

Figure 2: Approximated logarithmic B-spline functions. (a) Input exponential decay kernel functions (b) Learned 4-order polynomial mapping without a bias term.

Dataset	Representation	Ch.	AP 50:95	AP 50	AP 75	AP s	AP m	AP l	AR s
PEDRo	CSTR [ 7 ]	3	0.566	0.877	0.648	0.010	0.563	0.597	0.072
	Exp. decay	12	0.621	0.907	0.739	0.015	0.593	0.674	0.142
	Exp. decay + local conf.	12	0.621	0.908	0.734	0.018	0.600	0.668	0.106
	Spline	12	0.625	0.907	0.738	0.041	0.606	0.671	0.151
	Spline + local conf.	12	0.630	0.915	0.752	0.062	0.611	0.670	0.219
Gen1	CSTR [ 7 ]	3	0.355	0.637	0.349	0.278	0.435	0.274	0.446

Table 1: Comparison of temporal representations on PEDRo and Gen1. PEDRo models use a 40 ms event window, while Gen1 models use a 50 ms window. Only the event representation is changed within each dataset. Four temporal functions are used for both exponential and spline-based representations.

Panel of Figure 3 — Figure 3: Qualitative comparison of CSTR, exponential decay, and logarithmic spline representations on PEDRo. Each row shows one test example with predictions (red), annotations (green), and detector heatmaps. Exponential and spline representations are averaged over four temporal scales; red/green encode polarity support and blue local temporal confidence.

Window size (ms)	$\tau_{0}$ (ms)	$\tau_{\max}$ (ms)	AP 50:95	AP 50	AP 75	AP s
50	1.2	35	0.365	0.646	0.352	0.282
150	1.2	75	0.387	0.678	0.386	0.318
300	1.2	90	0.389	0.673	0.392	0.324
150	1.2	75	0.387	0.678	0.386	0.318
150	0.6	75	0.406	0.691	0.409	0.335
150	0.45	75	0.399	0.680	0.402	0.321

Table 2: Influence of window size and temporal function placement on Gen1. All models use four second-order logarithmic B-spline functions with local confidence support. All models were trained for 20 epochs using the same training configuration.

Window size (ms)	$\tau_{0}$ (ms)	$\tau_{\max}$ (ms)	AP 50:95	AP 50	AP 75	AP s
40	0.6	20	0.633	0.913	0.747	0.024
40	0.6	35	0.627	0.906	0.750	0.002
40	1.2	20	0.620	0.904	0.733	0.021
40	1.2	35	0.630	0.915	0.752	0.062
40	2.4	20	0.631	0.911	0.745	0.032
40	2.4	35	0.628	0.910	0.747	0.025

Table 3: Influence of temporal function placement on PEDRo. All models use four second-order logarithmic B-spline functions and local confidence support. All models were trained for 120 epochs using the same training configuration.

Panel of Figure 4 — Figure 4: Qualitative visualization of the logarithmic spline representation on Gen1 using a 150 ms event window. Columns show the averaged representation with predictions (red) and annotations (green), the detector heatmap, and the four spline channels spanning 0.6–75 ms. Red/green denote confidence-weighted polarity support and blue denotes local temporal confidence.

Method	Coeff.	MSE $\downarrow$	AP 50:95	AP 50	AP 75	AP s	AP m	AP l
Analytical spline	Exact	–	0.636	0.918	0.752	0.111	0.618	0.674
2nd order approx.	9	$1.44{\times}10^{-5}$	0.624	0.914	0.739	0.091	0.603	0.665
4th order approx.	34	$\mathbf{1.03{\times}10^{-6}}$	0.626	0.913	0.739	0.087	0.607	0.667

Table 4: Performance of recursive exponential-polynomial approximation of the logarithmic B-spline representation on PEDRo.

Dataset	Model	Recurrent.	Temporal context	Params	AP 50	AP 50:95
PEDRo	EventCenterNet (Ours)	No	40 ms	25M	0.918	0.636
	YOLOv8x [ 25 ]	No	40 ms	58M	0.895	0.586
	ReYOLOv8n [ 25 ]	Yes	200 (40 $\times$ 5) ms	4.7M	0.878	0.639
	ReYOLOv8s [ 25 ]	Yes	200 (40 $\times$ 5) ms	8.3M	0.905	0.655
	ReYOLOv8m [ 25 ]	Yes	200 (40 $\times$ 5) ms	18.1M	0.908	0.692
Gen1	EventCenterNet (Ours)	No	150 ms	25M	0.691	0.406

Table 5: Comparison with feed-forward and recurrent YOLO-based detectors. ReYOLO AP 50 and AP 50:95 values were obtained by re-evaluating the released models using the original evaluation protocol [ 25 ] , reproducing the published AP 50:95 results.

Method	AP 50:95	AP 50	AP 75
Spline + local conf.	0.636	0.918	0.752
+ Kalman, center + size	0.635	0.911	0.752
+ Kalman, size only	0.639	0.918	0.757

Table 6: Kalman-based temporal filtering on PEDRo.

Event-Based Vision

Event-based vision uses asynchronous camera measurements to perceive motion and scene changes with high temporal resolution and dynamic range.

See all 232 papers

Rethinking Event-Based Object Dtection through Representation-Level Temporal Aggregation and Model-Level Hypergraph Reasoning
May 9, 2026Meisen Wang, Hao Deng, Wei Bao +5Event-Based Vision Event-Based Object Detection
1Xi’an Jiaotong University · 2Tsinghua University · 3China Mobile System Integration +1
FATE: Pillar Encoding and Frequency-Aware Training for Event-Based Object Detection
Jun 15, 2026Md Tawheedul Islam Bhuian, Kyoung-Don KangEvent-Based Vision Frequency-Domain Feature Learning
School of Computing State University of New York at Binghamton
Neural Events: Discrete Asynchronous Autoencoders for Event-Based Vision
Jun 18, 2026Roberto Pellerito, Daniel Gehrig, Shintaro Shiba +1Event-Based Vision Representation Learning
Robotics and Perception Group, University of Zurich · University of Pennsylvania · The University of Tokyo +1

May 9, 2026J/K move · Enter open · S save

Rethinking Event-Based Object Dtection through Representation-Level Temporal Aggregation and Model-Level Hypergraph Reasoning

Meisen Wang, Hao Deng, Wei Bao, Ma Yuanxiao+4

1Xi’an Jiaotong University · 2Tsinghua University · 3China Mobile System Integration+1

Event-Based Vision Event-Based Object Detection Real-Time Object Detection

Open paper AskPDF arXiv

Event cameras provide microsecond-level temporal resolution, low latency, and high dynamic range, offering potential for perception under fast motion and challenging illumination conditions. However, existing Event-based Object Detection (EOD) methods face limitations at both the representation and model levels: prior event representations usually encode temporal information indirectly through redundant structures, while detection models struggle to explicitly aggregate fragmented event responses into coherent high-order object features. To address these limitations, we present Event Dual Temporal-Relational Aggregation Detector (Ev-DTAD), a unified EOD framework that integrates representation-level temporal encoding with model-level temporal-hypergraph reasoning. Specifically, we introduce Hierarchical Temporal Aggregation (HTA), a compact three-channel pseudo-RGB representation that explicitly embeds temporal information across intra- and inter-window events. To further enhance detection under sparse and fragmented event responses, we propose Frequency-aware Hypergraph Temporal Fusion (FHTF), which refines multi-scale event features through temporal evolution modeling and high-order relational reasoning. Extensive experiments on Gen1 (+0.8 mAP and 1.7

\times

faster), 1Mpx/Gen4 (+0.5 mAP and 1.6

\times

faster), and eTraM (+3.0 mAP and 2.0

\times

faster) demonstrate that Ev-DTAD achieves a competitive accuracy-efficiency trade-off, validating the complementarity between compact temporal representation and temporal-hypergraph feature reasoning.

Abstract

Event-Based Vision

Explore similar work

Rethinking Event-Based Object Dtection through Representation-Level Temporal Aggregation and Model-Level Hypergraph Reasoning

FATE: Pillar Encoding and Frequency-Aware Training for Event-Based Object Detection

Neural Events: Discrete Asynchronous Autoencoders for Event-Based Vision