Polarity--Time Interaction and Multi-Level Reliability for Event-Based Spacecraft Pose Estimation

Yonglong Zhang; Zongwu Xie; Yang Liu

cs.AIOct 27, 2026

Polarity--Time Interaction and Multi-Level Reliability for Event-Based Spacecraft Pose Estimation

Authors: Yonglong Zhang, Zongwu Xie, Yang Liu

Abstract

Sparse event responses produce uneven keypoint errors whose effects on spacecraft pose depend on correspondence geometry and the reliability of successive observations. We present a reliability-aware framework connecting polarity--time localization, geometric recovery, event--CAD refinement, and trajectory optimization. A three-branch representation combines positive events, negative events, and recent-event timing through cross-branch attention and gating, followed by spatiotemporal aggregation. An ordered semantic-point head estimates relative correspondence precision for geometrically constrained pose recovery. Event-edge and CAD-projection agreement then guide local pose refinement under keypoint-geometry safeguards. At the sequence level, a ten-dimensional descriptor combines heatmap shape, reprojection consistency, and geometric conditioning to set time-aligned observation scales. On the official SPADES challenge test set, the complete system achieves a combined pose score of 0.0474, approximately 5.4% below PASAN+IRIS. Validation ablations reduce the score from 0.0772 for base localization with uniform PnP to 0.0514 for the complete cascade; on identical single-frame observations, quality weighting reduces the score from 0.0549 with constant-weight optimization to 0.0514.

Figures & tables

Fig. 1: Event-based spacecraft pose estimation from detection to trajectory recovery. RTMDet-tiny detects the target, local harmonic fitting smooths the bounding-box trajectory, and the resulting boxes define the five-window ROI sequence. RPI establishes semantic correspondences and recovers an initial pose. ECR refines the pose using event–CAD agreement under geometric safeguards. MQKS combines the resulting observations with time-aligned quality and motion constraints to estimate the trajectory. Solid arrows denote the estimation flow and geometric inputs; the dashed arrow carries auxiliary quality information.

Fig. 2: Reliability-aware pose initialization. (A) Positive counts, negative counts, and recent-event timing interact through cross-branch attention and channel gates. The projected interaction residual is added to the polarity representation before five-window STA and spatial feature extraction. (B) Decoded semantic coordinates and local heatmap responses provide geometric and response-shape descriptors for point reliability. Predicted relative precision weights the geometric pose refinement.

Method / reported team	$E_{t}$	$E_{R}$	$E_{\mathrm{pose}}$
PASAN+IRIS	0.0146	0.0355	0.0501
RF-DETR cascade	0.0230	0.0331	0.0561
haoranhuang	0.1464	0.5562	0.7026
pruthj	0.1335	0.9101	1.0436
csu nuaa pang	0.3340	0.7724	1.1064
zhaoyutao22	0.1959	1.0607	1.2565

TABLE I: Comparison on the official SPADES challenge test set [ 3 , 4 ] . Lower scores indicate better performance.

Configuration	$E_{t}$	$E_{R}$	$E_{\mathrm{pose}}$
Base localization + uniform PnP	0.0208	0.0564	0.0772
RPI	0.0191	0.0506	0.0697
RPI + ECR	0.0185	0.0489	0.0674
RPI + MQKS	0.0154	0.0375	0.0529
RPI + ECR + MQKS	0.0150	0.0364	0.0514

TABLE II: Stage-wise ablation on the validation set.

Event representation	Fusion	Geometric weighting	$E_{t}$	$E_{R}$	$E_{\mathrm{pose}}$
Positive / negative counts	Dual-branch fusion	Uniform	0.0208	0.0564	0.0772
Counts + recent-event time	Concatenation + projection	Uniform	0.0204	0.0558	0.0762
Counts + recent-event time	Cross-branch attention	Uniform	0.0200	0.0539	0.0739
Counts + recent-event time	Attention + gating	Uniform	0.0197	0.0532	0.0729
Counts + recent-event time	Attention + gating	Learned, NLL	0.0195	0.0521	0.0716
Counts + recent-event time	Attention + gating	Learned, NLL + pose proxy	0.0191	0.0506	0.0697

TABLE III: Front-end design ablation with STA fixed and ECR and MQKS disabled. NLL denotes the localization-error likelihood objective.

Configuration	$E_{t}$	$E_{R}$	$E_{\mathrm{pose}}$
Single-frame poses	0.0185	0.0489	0.0674
Constant-weight optimization	0.0158	0.0391	0.0549
Heatmap-quality optimization	0.0156	0.0380	0.0536
Geometry-quality optimization	0.0152	0.0370	0.0522
MQKS	0.0150	0.0364	0.0514

TABLE IV: Quality-source ablation on identical single-frame observations.

Windows	$E_{t}$	$E_{R}$	$E_{\mathrm{pose}}$	ms/frame
1	0.0194	0.0483	0.0677	36.8
3	0.0162	0.0392	0.0554	44.6
5	0.0150	0.0364	0.0514	52.7
7	0.0149	0.0366	0.0515	61.3
9	0.0151	0.0372	0.0523	70.2

TABLE V: Effect of input window length on the proposed system.

Fig. 3: Full-system pose estimates on three preselected validation sequences, using seed 2026. Rows denote sequences and columns denote fixed timestamps. Green dashed and magenta solid lines show the CAD bounding box projected with the ground-truth and final estimated poses, respectively. Positive and negative events are orange and blue. All panels use a $448\times 448$ display grid covering 1.5 times the original ROI side length.

Fig. 4: Single-frame front-end pose estimates after ECR and before MQKS temporal optimization. Rows show three example sequences, and columns show four timestamps. Orange and blue pixels represent positive- and negative-polarity events, respectively. Red, green, and blue line segments denote the projected object-frame $X$ , $Y$ , and $Z$ axes, with a common physical length across all panels.

Fig. 5: Continuous-sequence errors for the three examples in Figure 3 , using seed 2026. Left: absolute translation error in meters. Right: rotation error in degrees. Gray curves show the single-frame ECR observations; blue dashed and orange solid curves show constant-weight optimization and MQKS, respectively. All 599 samples and error peaks are retained, without additional smoothing for visualization.

Spacecraft Pose Estimation

Spacecraft pose estimation recovers a spacecraft’s position and orientation from images and other sensor data for navigation, rendezvous, and servicing.

See all 19 papers

Efficient Onboard Spacecraft Pose Estimation with Event Cameras and Neuromorphic Hardware
Apr 5, 2026Arunkumar Rathinam, Jules Lecomte, Jost Reelsen +3Spacecraft Pose Estimation Relative Pose Estimation
University of Luxembourg · Fortiss GmbH, Munich, Germany · Technical University of Munich, Munich, Germany +1
Recurrent Vision Transformers for Event-Based Spacecraft Pose Estimation Under Slow Relative Motions
Oct 19, 2026Sofia McLeod, Mohsi Jawaid, Carl Seubert +1Spacecraft Pose Estimation Event-Based Vision
AI for Space Group, Adelaide University · Australasian Space Innovation Institute
Spacecraft Pose Tracking via Kinematic Prediction and Selective Event Correction
Oct 14, 2026Yonglong Zhang, Guanning Yao, Shuo Zhang +3Spacecraft Pose Estimation Event-Based Vision
All authors are with the State Key Laboratory of Robotics and Systems, Harbin Institute of Technology, Harbin 150001, Heilongjiang, China.

Apr 5, 2026J/K move · Enter open · S save

Efficient Onboard Spacecraft Pose Estimation with Event Cameras and Neuromorphic Hardware

Arunkumar Rathinam, Jules Lecomte, Jost Reelsen, Gregor Lenz+2

University of Luxembourg · Fortiss GmbH, Munich, Germany · Technical University of Munich, Munich, Germany+1

Spacecraft Pose Estimation Relative Pose Estimation Event-Based Vision Neuromorphic Computing

Open paper AskPDF arXiv

Reliable relative pose estimation is a key enabler for autonomous rendezvous and proximity operations, yet space imagery is notoriously challenging due to extreme illumination, high contrast, and fast target motion. Event cameras provide asynchronous, change-driven measurements that can remain informative when frame-based imagery saturates or blurs, while neuromorphic processors can exploit sparse activations for low-latency, energy-efficient inferences. This paper presents a spacecraft 6-DoF pose-estimation pipeline that couples event-based vision with the BrainChip Akida neuromorphic processor. Using the SPADES dataset, we train compact MobileNet-style keypoint regression networks on lightweight event-frame representations, apply quantization-aware training (8/4-bit), and convert the models to Akida-compatible spiking neural networks. We benchmark three event representations and demonstrate real-time, low-power inference on Akida V1 hardware. We additionally design a heatmap-based model targeting Akida V2 and evaluate it on Akida Cloud, yielding improved pose accuracy. To our knowledge, this is the first end-to-end demonstration of spacecraft pose estimation running on Akida hardware, highlighting a practical route to low-latency, low-power perception for future autonomous space missions.

Abstract

Spacecraft Pose Estimation

Explore similar work

Efficient Onboard Spacecraft Pose Estimation with Event Cameras and Neuromorphic Hardware

Recurrent Vision Transformers for Event-Based Spacecraft Pose Estimation Under Slow Relative Motions

Spacecraft Pose Tracking via Kinematic Prediction and Selective Event Correction