TL;DR
This paper presents a comparative study of two distinct approaches for event-based spacecraft pose estimation: a modular geometry-based pipeline and an end-to-end direct regression system. The authors uncover a surprising "performance reversal" where direct regression excels at orientation while the modular approach is far superior for translation, leading to a proposed hybrid architecture for future space missions.
What Problem Does This Paper Address?
Autonomous spacecraft operations like in-orbit docking or debris removal require precise 6-DoF (degrees of freedom) pose estimation—knowing exactly where a target is and how it's oriented. While traditional cameras are the standard, they often fail in the harsh lighting and high-speed motion of space. Event cameras, which report per-pixel brightness changes asynchronously, offer a promising alternative with high dynamic range and low latency.
However, a massive "sim-to-real" domain gap exists. Because real orbital data is scarce, models are often trained exclusively on synthetic data. When tested on real hardware, these models' performance typically declines due to sensor noise, dead pixels, and background artifacts. This paper investigates how two different system designs—one based on geometric recovery and one on direct neural regression—handle this gap and distribute their errors.
Methods
The authors compared two independently developed systems:
- Direct Regression System: This end-to-end approach uses two separate CNN backbones (EfficientNet-V2 and ConvNeXt). A translation branch processes a global view of the event stream, aided by an "apparent-scale" feature derived from the event cloud's spread. A rotation branch focuses on an adaptive, satellite-centered crop to capture fine structural details. To bridge the domain gap, it uses a suite of physics-motivated synthetic corruptions during training.
- Modular Keypoint-Tracking System: This pipeline decomposes the problem. First, an HRNet backbone predicts 2D keypoints on the spacecraft. Then, a geometric solver (EPnP with RANSAC) recovers the 3D pose by matching these 2D points to a known 3D model. It uses advanced domain adaptation techniques like CycleGAN and DANN to align synthetic and real data distributions.
Datasets
The study utilizes the SPADES (Spacecraft Pose Estimation Dataset) benchmark, which provides paired synthetic and real-world hardware recordings of an ESA PROBA-2 satellite mockup. Evaluation was conducted using the SPARK 2026 challenge hidden-test protocol, which requires predictions for all test frames.
Key Insights and Ideas
- The "Performance Reversal": The most striking finding is that the two systems have complementary strengths. The direct regression system is significantly better at estimating orientation, while the modular system is vastly superior at estimating translation.
- Why the Split? The authors hypothesize that modular systems benefit from the hard constraints of 3D geometry for translation. In contrast, direct regression's translation estimate relies on global scale cues that can be easily "inflated" by background noise. For rotation, however, the direct system's ability to learn robust features from cropped views outperforms the modular system's reliance on individual keypoints, which can be easily misplaced by noise.
- Metric Masking: Aggregate challenge scores can be misleading. Because rotation error (in radians) and relative translation error (unitless) have different numerical scales, the aggregate score in this benchmark favored the direct system, masking its significantly higher translation error.
Key Results
- Orientation: The direct regression system reduced orientation error by 22.7% (from 1.1767 to 0.9101 rad) compared to the modular system.
- Translation: The modular system was the clear winner for positioning, reducing relative translation error by a factor of 4.7, or 78.7% (from 0.1335 to 0.0285).
- Overall Score: The direct system achieved a 13.4% lower composite score, primarily because the metric was dominated by rotation performance.
Limitations and Future Work
The authors acknowledge that because the two systems differ in many ways—including neural backbones, data access, and temporal processing—the observed reversal cannot be attributed to a single design choice in isolation. For instance, the modular system used unlabeled target-domain data for adaptation, while the direct system relied on synthetic corruptions.
The findings motivate a new hybrid design hypothesis: combining direct orientation estimation with geometry-constrained translation recovery. Future work will involve implementing this hybrid and evaluating its components through controlled ablations across additional spacecraft and sensing conditions.
Summary updated