Locally Adaptive Decay Surfaces for High-Speed Face and Landmark Detection with Event Cameras

Paul Kielty; Timothy Hanley; Peter Corcoran

cs.CVFeb 26, 2026

Locally Adaptive Decay Surfaces for High-Speed Face and Landmark Detection with Event Cameras

Authors: Paul Kielty, Timothy Hanley, Peter Corcoran

Organizations: With the College of Science and Engineering, University of Galway, Ireland.

Abstract

Event cameras record luminance changes with microsecond resolution, but converting their sparse, asynchronous output into dense tensors that neural networks can exploit remains a core challenge. Conventional histograms or globally-decayed time-surface representations apply fixed temporal parameters across the entire image plane, which in practice creates a trade-off between preserving spatial structure during still periods and retaining sharp edges during rapid motion. We introduce Locally Adaptive Decay Surfaces (LADS), a family of event representations in which the temporal decay at each location is modulated according to local signal dynamics. Three strategies are explored, based on event rate, Laplacian-of-Gaussian response, and high-frequency spectral energy. These adaptive schemes preserve detail in quiescent regions while reducing blur in regions of dense activity. Extensive experiments on the public data show that LADS consistently improves both face detection and facial landmark accuracy compared to standard non-adaptive representations. At 30 Hz, LADS achieves higher detection accuracy and lower landmark error than either baseline, and at 240 Hz it mitigates the accuracy decline typically observed at higher frequencies, sustaining 2.44 % normalized mean error for landmarks and 0.966 mAP50 in face detection. These high-frequency results even surpass the accuracy reported in prior works operating at 30 Hz, setting new benchmarks for event-based face analysis. Moreover, by preserving spatial structure at the representation stage, LADS supports the use of much lighter network architectures while still retaining real-time performance. These results highlight the importance of context-aware temporal integration for neuromorphic vision and point toward real-time, high-frequency human-computer interaction systems that exploit the unique advantages of event cameras.

Figures & tables

Fig. 1: Visualization of the LADS process described in Section III-A . The input window of events, $\bm{\mathcal{E}}_{k}$ , is accumulated into histogram $H_{k}$ . The subsequent annotated steps are: (a) divide the histogram into a user-specified number of patches; (b) measure the signal dynamics of patch according to the selected LADS variant; (c) compute a decay value for each patch, which are interpolated to generate the per-pixel decay map $d_{k}$ ; (d) decay the previous surface, $S_{k-1}$ , by multiplication with $d_{k}$ (element-wise). The resulting product is added to $H_{k}$ to construct the new surface, $S_{k}$ .

Panel of Figure 2 — Fig. 2: Sample representations from an event video showing a blink on an otherwise still face. In the FFT examples, grid lines mark patch boundaries and a heatmap generated from the decay value assigned to each patch (before interpolation). With fewer patches, the recursive approach reduces computation while maintaining precise localization of the blink motion, which helps prevent nearby stationary features, such as the eyebrows, from receiving elevated decay values.

Panel of Figure 3 — Fig. 3: Each row shows a different stage of an event video: (a) during rapid head motion, (b) minimal head motion sustained for 0.5 s, and (c) minimal head motion sustained for over 2 s. Columns, left to right, are: histogram representation; time-surface with standard global LI; and time-surfaces generated by the three proposed adaptive integration methods (ER, LoG, FFT).

Panel of Figure 4 — Fig. 4: Examples of exclusions from the FES dataset with correct elements in green and incorrect elements in red: (a) Inconsistent landmark indexing. (b) Correct bounding box but incorrect landmark positions. (c) Incorrect bounding box but correct landmark positions.

Partition	# Subjects	Minutes of Video	# Samples
Train	39	283	509k
Valid	12	79.2	143k
Test	12	75.4	136k
All	63	437.6	788k

TABLE I: FES dataset partition details after cleaning

Representation	Face Detection mAP 50		Landmark NME (%)
	30Hz	240Hz	30Hz	240Hz
Histogram	0.921	0.829	2.60	4.05
Global-LI	0.948	0.929	2.37	2.76
ER	0.955	0.942	2.38	2.55
LoG	0.957	0.943	2.29	2.52
FFT	0.950	0.932	2.30	2.53

TABLE II: Comparison of different event representations for face and landmark detection on FES dataset at 30Hz and 240Hz

Representation	Face Detection mAP 50
	30Hz	240Hz
Histogram	0.394	0.186
Global-LI	0.552	0.536
ER	0.778	0.872
LoG	0.881	0.896
FFT	0.741	0.809

TABLE III: Comparison of different event representations for face detection on Blink dataset at 30Hz and 240Hz

Panel of Figure 8 — Fig. 5: LADS-ER representation of an event window featuring fast head motion from FES and Blink datasets at 30 Hz and 240 Hz.

Representation	Time (ms)
	30Hz	240Hz
Histogram	0.50	0.24
Global-LI	0.59	0.28
ER	1.19	0.60
LoG	0.95	0.52
FFT	1.86	0.61

TABLE IV: Mean time to construct the different event representations for each frame of the video shown in Fig. 3 at 30Hz and 240Hz. System hardware: NVIDIA GeForce RTX 2080 Ti, Intel Core i9-9960X @ 3.10 GHz, 128 GiB DDR4 @ 3 GHz

Frequency	Model	Event Input	# Params (M)	Face mAP 50	Landmark NME (%)
20 Hz	[ 17 ]	Histogram	24.1	0.973	2.44
20 Hz	[ 19 ]	Event Spikes	1.2	0.581	-
30 Hz	[ 17 ]	Histogram	24.1	0.370	6.35
	[ 19 ]	Event Spikes	1.2	0.557	-
	[ 4 ]	Global-LI	9.2	-	4.68
	Ours	LADS-LoG	2.6	0.969	-

TABLE V: Parameter counts and performance of event based face detectors (by mAP 50 ) and landmark detectors (by NME) of increasing inference frequencies, trained and tested on FES data.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

	Global-LI	ER		LoG		FFT
	$\tau$	$\tau$	$\lambda_{0}$	$\tau$	$a$	$r$	$T_{d}$
FES 30Hz	0.05	0.05	16	12.5	0.25	0.25	0.5
FES 240Hz	0.05	0.05	16	12.5	0.25	0.05	0.5
Blink 30Hz	0.2	0.2	2	7.5	0.75	0.01	0.9
Blink 240Hz	0.2	0.2	2	7.5	0.75	0.01	0.9

Appendix

TABLE VI: Parameters for all decay strategies and datasets at 30Hz and 240Hz.

Explore similar work

Apr 8, 2026

EventFace: Event-Based Face Recognition via Structure-Driven Spatiotemporal Modeling

Event cameras offer a promising sensing modality for face recognition due to their inherent advantages in illumination robustness and privacy-friendliness. However, because event streams lack the stable photometric appearance relied upon by conventional RGB-based face recognition systems, we argue that event-based face recognition should model structure-driven spatiotemporal identity representations shaped by rigid facial motion and individual facial geometry. Since dedicated datasets for event-based face recognition remain lacking, we construct EFace, a small-scale event-based face dataset captured under rigid facial motion. To learn effectively from this limited event data, we further propose EventFace, a framework for event-based face recognition that integrates spatial structure and temporal dynamics for identity modeling. Specifically, we employ Low-Rank Adaptation (LoRA) to transfer structural facial priors from pretrained RGB face models to the event domain, thereby establishing a reliable spatial basis for identity modeling. Building on this foundation, we further introduce a Motion Prompt Encoder (MPE) to explicitly encode temporal features and a Spatiotemporal Modulator (STM) to fuse them with spatial features, thereby enhancing the representation of identity-relevant event patterns. Extensive experiments demonstrate that EventFace achieves the best performance among the evaluated baselines, with a Rank-1 identification rate of 94.19% and an equal error rate (EER) of 5.35%. Results further indicate that EventFace exhibits stronger robustness under degraded illumination than the competing methods. In addition, the learned representations exhibit reduced template reconstructability.

Qingguo Meng, Xingbo Dong, Zhe Jin +1

State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, the Anhui Provincial Key Laboratory of Secure Artificial Intelligence, Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, and the School of Artificial Intelligence, Anhui University, Hefei, China · Computer Vision Laboratory, University of Sassari, 07100 Sassari, Italy

Face Recognition Event-Based Vision

Ask PDF

Mar 6, 2026

Low-Latency Event-Based Object Detection with Spatially-Sparse Linear Attention

Event cameras provide sequential visual data with spatial sparsity and high temporal resolution, making them attractive for low-latency object detection. Existing asynchronous event-based neural networks exploit this low-latency advantage by updating predictions event by event, but still suffer from two bottlenecks: recurrent architectures are difficult to train efficiently on long sequences, and improving accuracy often increases per-event computation and latency. Linear attention is appealing because it enables parallel training and recurrent inference. However, its dense state updates make per-event computation scale with the state size, yielding a poor accuracy-efficiency trade-off for object detection, where accurate localization requires fine-grained spatial states. The key challenge is therefore to introduce sparse state activation that exploits the spatial sparsity of events while preserving efficient parallel training. We propose Spatially-Sparse Linear Attention (SSLA), which introduces a mixture-of-spaces state decomposition and a scatter-compute-gather training procedure, enabling state-level sparsity as well as training parallelism. Building on SSLA, we develop an end-to-end asynchronous linear attention model, SSLA-Det, for low-latency event-based object detection. On Gen1 and N-Caltech101, SSLA-Det achieves state-of-the-art accuracy among asynchronous methods, reaching 0.375 mAP and 0.515 mAP, respectively, while reducing per-event computation by over 20 times compared with the strongest prior asynchronous baseline, demonstrating the potential of linear attention for low-latency event-based vision.

Haiqing Hao, Zhipeng Sui, Rong Zou +4

State Key Laboratory of Precision Measurement Technology and Instruments, Department of Precision Instrument, Tsinghua University, Beijing, China · Robotics and Perception Group, University of Zurich, Zurich, Switzerland · ShanghaiTech University, Shanghai, China

Efficient Object Detection Linear Attention Event-Based Object Detection Real-Time Object Detection

Ask PDF

Sep 17, 2026

An Event Preserving Velocity Invariant Representation for Event Cameras

Event cameras provide low-latency, high temporal resolution perception for real-time vision tasks such as robotics.The novel circuitry (i.e. asynchronous, independent pixels) that enables these advantages also introduces new algorithmic challenges. Velocity-invariant representations alleviate missing observations under slow motion and motion blur under fast motion, but most discard temporal information by converting events into image-like representations. We propose Set of Centre Active Receptive Fields (SCARF), a real-time velocity-invariant representation that preserves raw events while consistently handling fast motion, stationary scenes, and independently moving objects. SCARF achieves state-of-the-art performance in both computational efficiency and representation quality.

Mikihiro Ikura, Luna Gava, Jiahang Wu +2

Istituto Italiano di Tecnologia, Via Morego 30 16163 Genova, Italy

Event-Based Vision

Ask PDF