cs.CVAug 27, 2026

Real-time Unsupervised Object Discovery from Asynchronous Event Streams

Authors: Pratham G. Shenwai, Hemant Kumar Singh, Sridhar Ravi

Organizations: School of Engineering and Technology University of New South Wales Canberra, Australia

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Event cameras capture pixel-level intensity changes with microsecond resolution to produce highly sparse asynchronous data streams. For visual perception in latency-critical environments, we propose a lightweight, training-free framework for discovery of moving objects based on spatio-temporal clustering. This framework is driven by two core contributions. First, a linear-time Spatio-temporal Probabilistic Event Filter (SPEF) that introduces an adaptive event acceptance threshold to distinguish salient motion structures from background noise. Second, an Event Morton Code Clustering (EMCC) module that bypasses expensive distance matrix computation to efficiently group events for unsupervised discovery of moving objects. On the E-MLB dataset benchmark, SPEF achieves the best denoising performance among classical filtering methods and remains competitive with learning-based approaches without requiring any offline training. On object discovery, EMCC achieves the highest overall accuracy and lowest execution time across the FRED and eTraM datasets, outperforming established density-based clustering baselines by a substantial margin. Overall, this work establishes a new performance benchmark for classical object discovery in event data, providing a highly scalable, training-free solution for resource-constrained visual perception. The code is available at https://github.com/PrathamShenwai/SPEF_EMCC

Figures & tables

Figure 1: Overview of the proposed unsupervised event - based object discovery framework. Events are processed in sliding temporal windows, denoised using the proposed Spatio - temporal Probabilistic Event Filter (SPEF) , encoded with Morton codes to preserve spatio - temporal locality, and clustered via Event Morton Code Clustering (EMCC) to produce 2D bounding box proposals.
Figure 2: Overview of our object discovery framework using asynchronous event streams. The framework begins with the Spatio-temporal Probabilistic Event Filter (SPEF), which maps events to grid cells and updates correlation scores using inter-event timing and probabilistic counters. Active regions are isolated via a top-k% threshold, and events are stochastically filtered to maximize signal-to-noise while preserving structural integrity. The accepted events then flow into the Event Morton Code Clustering (EMCC) module. Here, they are normalized and assigned n-bit Morton (Z-order) codes to strictly preserve spatiotemporal locality. Sorting these codes allows for highly efficient clustering via gap detection in Morton space. Finally, candidate clusters are validated using geometric and statistical constraints to yield the final bounding box predictions.
Figure 3: Comparison of denoising methods on three E-MLB [ 81 ] sequences across ND04 and ND64 noise levels
 MethodE-MLB (Day)E-MLB (Night)
 ND1ND4ND16ND64ND1ND4ND16ND64
Model-basedRaw0.8210.8240.8150.7860.8900.8240.7860.768
BAF [ 88 ]0.8610.8690.8760.8900.9460.9730.9920.942
TS [ 89 ]0.8770.8870.8700.8371.0330.9440.8860.797
KNoise [ 19 ]0.8460.8370.8300.8070.9540.9560.8710.817
EvFlow [ 90 ]0.8480.8780.8680.8330.9690.9830.8890.797
Table 1: Comparison of event denoising methods using the MESR ↑\uparrow metric on E-MLB [ 81 ] dataset. Best results are bold , and second best are underlined .
Figure 4: Performance profiles of four clustering algorithms evaluated on the FRED [ 82 ] dataset. The curves plot the performance ratio threshold ( τ\tau ) against the fraction of sequences ( ρ(τ)\rho(\tau) ) where an algorithm’s metric is within a factor of τ\tau of the top-performing method. Steeper curves that reach ρ(τ)=1.0\rho(\tau)=1.0 at lower τ\tau values indicate greater performance consistency across the dataset. As shown in the profiles, ST-MeanShift (red dashed) achieves the highest initial Precision at τ=1\tau=1 . However, our proposed EMCC (purple dash-dotted) exhibits robust overall performance, reaching ρ(τ)=1.0\rho(\tau)=1.0 faster than ST-DBSCAN, ST-MeanShift, and ST-HDBSCAN in Recall, F1, and IOU. Notably, EMCC achieves the best relative performance ( ρ(τ)=1.0\rho(\tau)=1.0 ) for E2E Latency across all sequences at τ=1\tau=1 and shows significantly higher consistency in E2E Memory compared to the baselines. Profiles shown: (a) Precision, (b) Recall, (c) F1, (d) IOU, (e) Mean E2E Latency, and (f) Mean E2E Memory.
Figure 5: Qualitative comparison of object discovery on FRED [ 82 ] and eTraM [ 83 ] datasets. For FRED, yellow boxes on RGB frames provide visual reference, with black-dashed insets showing zoomed-in views of discovered objects. Bounding box proposals are displayed for ST-DBSCAN (blue), ST-MeanShift (red), ST-HDBSCAN (green), EMCC (purple-dashed), and Ground Truth (black).
MethodP ↑\uparrowR ↑\uparrowF1 ↑\uparrowIOU@50 ↑\uparrowE2E Lat. (ms) ↓\downarrow
FRED [ 82 ]
ST-DBSCAN [ 59 ]0.3480.4020.3580.59427.3
ST-Mean Shift [ 84 ]0.5110.2600.3010.631517.1
ST-HDBSCAN [ 60 ]0.2390.2810.2480.60270.4
EMCC (Ours)0.4160.5170.4410.6476.3
eTraM [ 83 ]
Table 2: Unsupervised Spatio-temporal Object Discovery Performance on FRED and eTraM. All methods process SPEF-filtered event streams. ST denotes Spatio-temporal and indicates our x,y,tx,y,t implementations of the referenced algorithms. Latency is reported in ms. Best results are bold , and second best are underlined .
SPEF
ParamConfigPRF1IoULat.
counter bits (b)counter~bits~(b)Raw (No Filter)0.3400.2490.2560.62314.6
 Probab. (5-bit)0.4270.4880.4340.6435.9
 Linear (5-bit)0.4160. 0.5170.4410.6476.3
 Probab. (2-bit)0.4160.3430.3530.6506.5
 Linear (2-bit)0.4220.3040.3190.6487.8
Table 3: Ablation results for the SPEF and EMCC modules. Bold marks the best value per group. Baseline settings are shaded.
HyperparametersFRED [ 82 ]eTraM [ 83 ]
ST-DBSCAN  
eps0.005–0.10.005–0.3
min samples3–503–50
ST-HDBSCAN  
min cluster size5–605–200
min samples1–301–30
Table ST1: Hyperparameter search ranges for the evaluated clustering algorithms on FRED [ 82 ] and eTraM [ 83 ] . Shared hyperparameters govern bounding box generation and are jointly optimised alongside method-specific parameters, with each algorithm receiving its own independently optimised shared parameter set.
Figure SF1: Comparison of probabilistic (RNG) and fixed-threshold acceptance across three representative scenarios: (a) multiple objects (Instance 63), (b) distant sparse target (Instance 90), and (c) dynamic proximity changes (Instance 154). Subplots (i) show acceptance probability over time and subplots (ii) report Precision, Recall, F1, and IoU@50. RNG (blue) adapts dynamically to scene activity, outperforming fixed thresholds at θfix∈{0.4,0.5,0.6,0.7,0.8}\theta_{\text{fix}}\in\{0.4,0.5,0.6,0.7,0.8\} across all scenarios.
Figure SF2: Qualitative comparison of object detection outputs on the FRED and eTraM datasets: additional samples. Yellow boxes on RGB frames of FRED provide visual reference, with black-dashed insets showing zoomed-in views of discovered objects. Bounding box proposals are displayed for ST-DBSCAN (blue), ST-MeanShift (red), ST-HDBSCAN (green), EMCC (purple-dashed), and Ground Truth (black). These supplementary examples extend the qualitative analysis presented in the main paper, demonstrating EMCC’s performance consistency across diverse scene configurations and dynamic activity levels.
More in this topic

Unsupervised Visual Concept Discovery

See all 13 papers

Explore similar work

CardsList
Feb 26, 2026

Motion-aware Event Suppression for Event Cameras

Event cameras report asynchronously per-pixel brightness changes with microsecond latency, encoding dynamic visual information as a sparse stream of events. However, their extreme temporal resolution floods perception systems with entangled events from ego-motion and independently moving objects (IMOs), which existing solutions fail to efficiently decouple, relying instead on prohibitive dense 3D reconstructions or limited hand-tuned filters. In this work, we introduce the first framework for Motion-aware Event Suppression, which learns to filter events triggered by IMOs and ego-motion in real time. Our model jointly segments IMOs in the current event stream while predicting their future motion, enabling anticipatory suppression of dynamic events before they occur. Our lightweight architecture achieves 173 Hz inference on consumer-grade GPUs with less than 1 GB of memory usage, outperforming previous state-of-the-art methods on the challenging EVIMO benchmark by 67% in segmentation accuracy while operating at a 53% higher inference rate. Moreover, we demonstrate significant benefits for downstream applications: our method accelerates Vision Transformer inference by 83% via token pruning and improves event-based visual odometry accuracy, reducing Absolute Trajectory Error (ATE) by 13%.

Roberto Pellerito, Nico Messikommer, Giovanni Cioffi +2

Robotics and Perception Group, University of Zurich, Switzerland

Event-Based Vision
AskPDF
Sep 17, 2026

PointEvent: Rethinking Event-based Tiny Object Detection via Serialized Motion Evidence Accumulation

Event cameras offer high temporal resolution and motion sensitivity for tiny UAV detection, yet distant targets generate sparse and fragmented events that are easily overwhelmed by clutter and ego-motion. Existing methods mainly rely on dense event representations or local sparse spatiotemporal modeling, resulting in redundant computation or fragmented modeling of motion continuity across distant asynchronous events. To address this limitation, we introduce serialized motion evidence accumulation, which treats motion continuity as an ordered evidence propagation process. Specifically, the same event stream is organized into locality-preserving spatiotemporal paths and chronology-preserving temporal paths through the latent complementary serializations. Based on this principle, we propose PointEvent, a lightweight event-wise state-space framework that alternates serialized scans across the complementary orders, progressively consolidating fragmented motion evidence beyond fixed local neighborhoods. A high-resolution event branch preserves fine-grained target responses, while compact context modulation suppresses interference. Experiments demonstrate that PointEvent achieves SOTA with the fewest parameters and fastest measured inference among the compared methods. Code: https://github.com/wzz-z/PointEvent

Zongze Wu, Baofeng Jia, Weiqi Yan +4

State key Lab of Extreme Environment Optoelectronic Dynamic Testing Technology and Instrument, Nanjing University of Science and Technology, China · Jiangsu Key Lab of Visual Sensing and Intelligent Perception, Nanjing University of Science and Technology, China · Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University, China

Small Object DetectionEfficient Object DetectionEvent-Based Object DetectionAerial Object Detection
AskPDF
Jun 18, 2026

Neural Events: Discrete Asynchronous Autoencoders for Event-Based Vision

Event cameras capture dynamic scenes with exceptional temporal fidelity by representing them as a continuous stream of microsecond resolution events. Each individual event, however, only carries minimal semantic value, merely signaling a localized brightness change. To derive meaningful signals, downstream algorithms need to quickly integrate cues from a potentially massive torrent of low-information events. Current architectures, however, are easily overwhelmed, struggling to balance capturing fine-grained temporal dynamics and maintaining a manageable data throughput. This paper proposes a framework to re-tokenize event streams into a small set of highly informative neural events, each representing a local spatio-temporal context window with a discrete learnable code. Every time this code flips, a neural event is triggered, yielding a highly compressed data stream. We demonstrate that, across object detection and classification, networks trained on neural events are on par or surpass the performance of state-of-the-art approaches while reducing the event rate by a factor of 2.0.

Roberto Pellerito, Daniel Gehrig, Shintaro Shiba +1

Robotics and Perception Group, University of Zurich · University of Pennsylvania · The University of Tokyo +1

Event-Based VisionRepresentation LearningAutoencodersEvent-Based Object Detection
AskPDF