StarPose: Cross-Domain Pose Inference

• StarPose is a versatile framework that maps observations to latent geometry across domains like human pose, spacecraft vision, and stellar-surface inference.
• It employs autoregressive diffusion, historical memory integration, and spatial-temporal physical guidance to resolve ambiguities and ensure temporal consistency.
• Applications range from precise 3D human pose estimation to starshade sensing and Bayesian stellar surface reconstruction, highlighting both its strengths and limitations.

StarPose is a polysemous research term whose meaning depends on domain. In the supplied literature, it appears both as the title of a specific monocular 3D human pose estimation method based on autoregressive diffusion and as a conceptual label for broader pose- or surface-inference systems in starshade sensing, spacecraft relative navigation, and stellar-surface reconstruction. Across these uses, the common pattern is an inverse problem in which observations are mapped to latent geometric state—sometimes a rigid 6-DoF pose, sometimes a 2D lateral offset, and sometimes a stellar orientation and surface map (Yang et al., 4 Aug 2025).

1. Terminological scope and conceptual core

In its narrowest sense, StarPose denotes the framework introduced in "StarPose: 3D Human Pose Estimation via Spatial-Temporal Autoregressive Diffusion" (Yang et al., 4 Aug 2025). That method treats monocular 3D human pose estimation as a temporally autoregressive diffusion problem, conditioned on 2D pose input, historical 3D predictions, and explicit spatial-temporal physical guidance.

In the broader supplied literature, "StarPose" is also used as an interpretive label rather than a formal method name. The starshade position-sensing work describes itself as a conceptual "2D StarPose prototype" for inferring the lateral position $(x,y)$ of a starshade from pupil-plane diffraction images (Chen et al., 2022). The stellar-surface mapping paper states that StarryStarryProcess is essentially a "StarPose" framework because it infers stellar inclination, projected obliquity, and statistical spot structure from transit and rotational light curves (Sagynbayeva et al., 30 Apr 2025). SPICE and PAStar are presented as forward models that would naturally sit inside a StarPose-type inversion system for stars with non-homogeneous surfaces (Jabłońska et al., 14 Nov 2025, Petralia et al., 2024). In spacecraft vision, several papers do not use the name explicitly but describe datasets or methods as "StarPose-style" because they support or instantiate monocular pose estimation of a known spacecraft from synthetic imagery (Velkei et al., 15 Jun 2025, Zhang et al., 2 Jul 2026, Park et al., 2024).

Domain	Inferred quantity	Representative paper
Human pose estimation	3D joint coordinates from monocular 2D pose sequences	(Yang et al., 4 Aug 2025)
Starshade sensing	2D lateral position $(x,y)$ in the telescope pupil frame	(Chen et al., 2022)
Spacecraft vision	6-DoF relative pose of a known target spacecraft	(Zhang et al., 2 Jul 2026)
Stellar-surface inference	Inclination, obliquity, and surface spot structure	(Sagynbayeva et al., 30 Apr 2025)

This domain dependence is essential. A common misconception would be to treat StarPose as a single standardized architecture or benchmark. The supplied papers instead show a family resemblance: a measurement-to-geometry pipeline, often paired with strong physical structure, synthetic data, or explicit generative modeling.

2. StarPose as autoregressive diffusion for monocular 3D human pose

In the explicit human-pose formulation, StarPose addresses the standard 2D-to-3D lifting problem. Given a video clip of length $f$, the 2D joint sequence is

$X \in \mathbb{R}^{f \times (J\cdot 2)},$

and the target is the 3D pose of the central frame,

$y \in \mathbb{R}^{1 \times (J \cdot 3)}.$

The method is designed to resolve depth ambiguity and improve temporal consistency relative to deterministic Transformer- and CNN-based regressors (Yang et al., 4 Aug 2025).

The overall pipeline contains four main components. First, a pre-trained and frozen MixSTE encoder produces contextual 2D features $f_{\text{2D}}$. Second, the Historical Pose Integration Module (HPIM) aggregates past 2D poses and previously predicted 3D poses into a historical embedding $f_{\text{his}}$. Third, the core Autoregressive Pose Conditional Diffusion module denoises 3D pose hypotheses over time. Fourth, a plug-and-play Spatial-Temporal Physical Guidance mechanism steers denoising toward anatomically plausible and temporally smooth solutions (Yang et al., 4 Aug 2025).

The forward noising process is defined by

$$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$

where $\mu$ is the mean of a Gaussian Mixture Model over training poses, $\epsilon$ is sampled from the same GMM, and $(x,y)$0 is a decreasing schedule. The reverse process is conditioned on both present 2D context and historical pose information: $(x,y)$1 This explicitly makes the sequence model autoregressive across video time: $(x,y)$2

The denoising model uses GraFormer, with three stacked GCN-attention blocks. Key implementation settings reported in the paper are $(x,y)$3 input frames, $(x,y)$4 joints, $(x,y)$5 hypotheses, $(x,y)$6 diffusion steps during training, $(x,y)$7 GMM components, and $(x,y)$8 for both the HPIM history length and the bone-variance window. Training uses Adam for 50 epochs, an initial learning rate of $(x,y)$9, exponential decay by a factor of $f$0 every 10 epochs, and batch size 1024 on a single RTX 4090 GPU (Yang et al., 4 Aug 2025).

Quantitatively, the reported Human3.6M results with CPN 2D input are MPJPE $f$1 mm and P-MPJPE $f$2 mm. With ground-truth 2D, MPJPE is $f$3 mm. On temporal metrics, StarPose reports MPJVE $f$4 mm/s and ACC-ERR $f$5 mm/s$f$6, substantially below the cited previous baselines. On MPI-INF-3DHP, the paper reports PCK $f$7, AUC $f$8, and MPJPE $f$9 mm (Yang et al., 4 Aug 2025).

3. Historical conditioning and physical guidance

HPIM is the principal mechanism by which StarPose incorporates temporal memory. For a target frame $X \in \mathbb{R}^{f \times (J\cdot 2)},$0, it uses past 2D poses

$X \in \mathbb{R}^{f \times (J\cdot 2)},$1

and past 3D predictions

$X \in \mathbb{R}^{f \times (J\cdot 2)},$2

Each modality is projected into a joint-wise embedding, processed by spatial transformer encoders, and organized into a Skeleton Integration Graph whose nodes contain both 2D and 3D joints over time. The graph includes within-frame skeletal edges, 2D-3D cross-dimensional edges, and temporal edges linking the same joint across consecutive frames. A GCN branch captures short-range structure, an attention branch captures long-range structure, and the final history representation is

$X \in \mathbb{R}^{f \times (J\cdot 2)},$3

Ablation results attribute a substantial portion of the total accuracy gain to HPIM: adding HPIM alone reduces MPJPE from $X \in \mathbb{R}^{f \times (J\cdot 2)},$4 mm to $X \in \mathbb{R}^{f \times (J\cdot 2)},$5 mm relative to a DiffPose-like baseline (Yang et al., 4 Aug 2025).

The second major element is the Spatial-Temporal Physical Guidance mechanism. It defines an energy

$X \in \mathbb{R}^{f \times (J\cdot 2)},$6

combining four constraints: 2D reprojection consistency, left-right skeletal symmetry, temporal bone-length variance, and differential sequence variation. At each reverse step, the estimate is adjusted via

$X \in \mathbb{R}^{f \times (J\cdot 2)},$7

The reported default weights are $X \in \mathbb{R}^{f \times (J\cdot 2)},$8, $X \in \mathbb{R}^{f \times (J\cdot 2)},$9, $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$0, and $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$1 (Yang et al., 4 Aug 2025).

The reprojection term uses a 2D reprojection function $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$2,

$y \in \mathbb{R}^{1 \times (J \cdot 3)}.$3

while the symmetry penalty compares predefined left-right bone pairs,

$y \in \mathbb{R}^{1 \times (J \cdot 3)}.$4

The bone-variance and temporal-smoothness terms enforce invariance of skeletal length over a sliding window and reduce frame-to-frame jitter. The paper emphasizes that STPG is plug-and-play and can be attached to other diffusion models. In the reported plug-in experiments, DiffPose improves from $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$5 to $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$6 mm MPJPE and D3DP from $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$7 to $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$8 mm, with only about $y \in \mathbb{R}^{1 \times (J \cdot 3)}.$9–$f_{\text{2D}}$0 FPS reduction (Yang et al., 4 Aug 2025).

A plausible implication is that, within the human-pose setting, StarPose is best understood less as diffusion alone than as a three-part synthesis of temporally autoregressive conditioning, graph-structured historical memory, and gradient-based physical regularization.

4. StarPose-style spacecraft pose estimation

In spacecraft vision, the supplied literature uses "StarPose-style" to denote monocular RGB-based estimation of the 6-DoF pose of a known, non-cooperative target spacecraft. The underlying state is the standard relative pose

$f_{\text{2D}}$1

with $f_{\text{2D}}$2 and $f_{\text{2D}}$3 (Velkei et al., 15 Jun 2025).

DLVS$f_{\text{2D}}$4-HST-V1 is presented as a synthetic data source directly aligned with what a StarPose-like spacecraft method would need, even though the paper does not explicitly mention StarPose by name (Velkei et al., 15 Jun 2025). The initial rigid-body release contains $f_{\text{2D}}$5 images, with planned extensions of $f_{\text{2D}}$6 articulated images and $f_{\text{2D}}$7 scenario/trajectory images, for a full planned dataset of $f_{\text{2D}}$8 images at $f_{\text{2D}}$9 resolution. Each EXR sample provides HDR RGB, a camera-frame normal map, per-pixel depth, a Cryptomatte-style semantic segmentation with 19 non-overlapping HST sub-part IDs, 37 predefined 3D keypoints and their 2D projections, and metadata including CameraTransform, TargetTransform, and celestial positions. Camera-target distances are typically $f_{\text{his}}$0–$f_{\text{his}}$1 m, and the orbital sampling covers 640 orbits generated from SPICE-driven HST ephemerides (Velkei et al., 15 Jun 2025).

The simulation pipeline combines Unreal Engine 5 for real-time planetary backgrounds and Houdini Karma for offline ray tracing. It includes MaterialX-based physically based BRDFs, randomized material aging, Earthshine, and image-based lighting from $f_{\text{his}}$2 HDR environment maps. The paper frames these choices as a working hypothesis for reducing the sim-to-real domain gap, but it does not yet report a quantitative domain-gap evaluation or baseline pose-estimation results on the dataset (Velkei et al., 15 Jun 2025). A common misconception would be to read DLVS$f_{\text{his}}$3-HST-V1 as a benchmark with established StarPose baselines; the paper is dataset-oriented and explicitly leaves those comparisons for future work.

At the method level, GAP-GDRNet is described as "essentially a StarPose-style system for spacecraft" within the geometry-guided direct regression family (Zhang et al., 2 Jul 2026). It takes a single RGB crop, camera intrinsics, and a 2D bounding box, and predicts the full 6D pose. Two additions are emphasized: an Attention-based Feature Refinement module before dense geometry prediction and Patch-level Geometric Self-Attention inside Patch-PnP. On the reported single-target synthetic spacecraft dataset, baseline GDR-Net with ResNet-34 yields $f_{\text{his}}$4, $f_{\text{his}}$5 m, and [email protected] m $f_{\text{his}}$6, while full GAP-GDRNet reports $f_{\text{his}}$7, $f_{\text{his}}$8 m, and [email protected] m $f_{\text{his}}$9. The model is reported at 47.75M parameters, 14.99 GFLOPs, and 27.8 ms latency, or about 35.97 FPS (Zhang et al., 2 Jul 2026).

SPNv3 occupies the same problem setting but emphasizes flight readiness and source-only sim-to-real robustness. It estimates spacecraft pose through heatmap-based keypoint detection followed by EP$$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$0P, using a ViTPose backbone and synthetic-only training on SPEED+ imagery (Park et al., 2024). For HIL evaluation, the paper reports calibration-aware pose metrics and shows that SPNv3-S, SPNv3-M, and SPNv3-B reach competitive results on lightbox and sunlamp without using HIL images for training. SPNv3-S at $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$1 runs at about 34 ms per image on Jetson Nano 4GB, which the paper notes is well above the update frequency of modern satellite navigation filters (Park et al., 2024).

Taken together, these spacecraft papers suggest that StarPose-style systems in orbital robotics are defined less by one named algorithm than by a recurring stack: synthetic supervision, known 3D geometry, 2D keypoints or dense geometric correspondences, geometric solvers or learned PnP, and deliberate efforts to reduce the sim-to-real gap.

5. StarPose as starshade sensing and closed-loop formation control

The starshade position-sensing paper uses the term in a narrower but explicit analogical sense: the current work can be seen as a pose estimation system—"StarPose"—for a starshade, where pose means the 2D lateral position $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$2 of the starshade in the telescope pupil frame (Chen et al., 2022). The measurement is not a natural image but a $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$3 pupil-plane intensity image containing the spot of Arago in out-of-band light. The location of that peak corresponds directly to the starshade’s lateral offset.

The method couples a lightweight CNN with simulation-based inference. The CNN receives a single-channel $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$4 image and regresses $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$5. The architecture has three convolutional layers with channel sizes $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$6, $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$7, and $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$8, each followed by max-pooling and ReLU, then two fully connected layers $$h_k = \mu + \sqrt{\alpha_k}(h_0 - \mu) + \sqrt{1 - \alpha_k}\,\epsilon,$$9. It has about $\mu$0 parameters, a stored model size of 1.6 MB, and a test-time cost of 5.3 MFLOPs per image. Training uses 160,000 simulated images with positions uniformly sampled from a $\mu$1 m square and peak SNR uniformly from 0.5 to 100 (Chen et al., 2022).

To obtain uncertainty and correct bias, the method then fits the joint density $\mu$2 with DELFI using a Gaussian Mixture Model,

$\mu$3

with $\mu$4 components trained multiple times and averaged to $\mu$5. Conditioning on a new CNN output produces a posterior over $\mu$6, with the covariance interpreted as uncertainty and the mode component center used as a corrected point estimate (Chen et al., 2022).

On simulated $\mu$7 images with peak SNR $\mu$8, the mean radial error is 6.3 cm, with standard deviation 4.8 cm and 99.7% of errors below 25 cm. Restricting to $\mu$9 m, the mean error is 2.8 cm and 99.7% of errors are below 10.5 cm. In Princeton Starshade Testbed experiments over 23,700 images with peak SNR 5–8, the mean radial error is 6.5 cm, the standard deviation is 4.0 cm, and 99.7% of errors are below 25 cm. Within $\epsilon$0 m, the 99.7% percentile error is 21.5 cm, still satisfying the stated 30 cm requirement (Chen et al., 2022).

The sensing module is also integrated into a closed-loop formation-flying controller. A pupil image is acquired every second, the CNN produces an estimate, a UKF fuses estimates with a relative-motion model, and an LQR computes thruster commands. In the reported testbed demonstration, initial misalignment is about 0.5 m in $\epsilon$1, peak SNR is about 8, and the system keeps misalignment within $\epsilon$2 m in both axes for the mission duration. Simulations comparing DELFI-derived measurement noise against a fixed 2 cm noise model report a 3% reduction in average fuel consumption and an 8% improvement in UKF position-estimation error (Chen et al., 2022).

This usage clarifies another important point of terminology. In some StarPose-like work, "pose" does not mean full rigid-body $\epsilon$3; it can denote only those degrees of freedom that are observable and operationally relevant.

6. StarPose-like stellar surface and orientation inference

In stellar applications, the supplied literature extends the idea of pose beyond rigid geometry to the orientation and structure of a stellar surface. StarryStarryProcess is described as "essentially a 'StarPose' framework" because it performs joint Bayesian inference over spot-population hyperparameters, stellar inclination $\epsilon$4, projected obliquity $\epsilon$5, and planetary transit geometry, using occultation light curves and rotational modulation (Sagynbayeva et al., 30 Apr 2025).

Its core representation is a spherical-harmonic surface map,

$\epsilon$6

with a Gaussian prior on the coefficient vector $\epsilon$7 induced by a spot-population model. The flux model is linear in the map coefficients,

$\epsilon$8

and after marginalizing $\epsilon$9, the method obtains a Gaussian-process likelihood over the observed light curve. The parameters are partitioned into spot hyperparameters $(x,y)$00, stellar parameters $(x,y)$01, and planet parameters $(x,y)$02. The paper emphasizes that transits break null-space degeneracies that are unavoidable when using rotational light curves alone, because spot-crossing anomalies localize active latitudes and constrain mutual geometry (Sagynbayeva et al., 30 Apr 2025).

For the real system TOI-3884, the reported posterior includes $(x,y)$03 deg, $(x,y)$04 deg, $(x,y)$05 d, and a near-polar mean spot latitude $(x,y)$06 deg. The inferred latitude PDF is bimodal at $(x,y)$07 because photometry cannot break the north-south flip degeneracy. The paper states explicitly that a system with $(x,y)$08 and one with $(x,y)$09, together with a map reflected across the stellar equator, produce identical light curves (Sagynbayeva et al., 30 Apr 2025). In this setting, "pose" therefore includes orientation but remains only partially identifiable.

PAStar and SPICE play a different but closely related role. Both are presented as forward models that could be embedded in a StarPose-type inversion framework. PAStar models stellar photospheric activity using a three-component atmosphere—quiet photosphere, spots, and faculae—and computes photometric and spectroscopic observables while accounting for stellar inclination, rigid rotation, limb darkening, and Doppler shifts (Petralia et al., 2024). Its disk-integrated flux is built by summing contributions from surface elements, with either a linear or a Claret four-coefficient limb-darkening law. The paper reports that PAStar can retrieve input surface inhomogeneity configurations through photometric or spectroscopic observations, validates the model against optical solar data, and compares it with SOAP on synthetic faculae-dominated solar cases (Petralia et al., 2024).

SPICE, by contrast, is an open-source Python package for generating high-resolution spectra and photometry from non-homogeneous stellar surfaces. It represents the star as a triangular mesh, evaluates angle-dependent specific intensities $(x,y)$10 on each facet, applies Doppler shifts from rotation, pulsation, or orbital motion, and integrates the visible contributions,

$(x,y)$11

The paper explicitly states that SPICE is not itself an inversion method, but rather the differentiable forward map one would place inside a StarPose-like system for inferring stellar surface structure and pose from photometric and spectroscopic time series (Jabłońska et al., 14 Nov 2025).

Across these stellar papers, the central reinterpretation is that "pose" includes quantities such as spin-axis inclination, projected obliquity, transit-chord geometry, pulsation phase, and surface texture. This usage is technically different from human or spacecraft pose, but the underlying inverse-problem structure is closely analogous.

7. Common structure, limits, and recurrent misconceptions

Across the supplied literature, StarPose and StarPose-like systems share three recurrent design elements. The first is physically structured latent state: human skeletons with bone symmetries, spacecraft with known 3D geometry and camera models, starshades with diffraction-governed pupil signatures, and stars with spherical-harmonic or mesh-based surfaces. The second is strong synthetic or simulated supervision: simulated 3D human pose diffusion trajectories, high-fidelity starshade diffraction images, synthetic spacecraft renders with keypoints and dense geometry, and forward-generated stellar spectra or light curves (Yang et al., 4 Aug 2025, Chen et al., 2022, Velkei et al., 15 Jun 2025, Jabłońska et al., 14 Nov 2025). The third is explicit treatment of uncertainty or ambiguity, whether through DELFI posteriors, diffusion sampling, multimodal MCMC posteriors, or calibration-aware pose metrics.

Several limitations also recur. Human StarPose depends strongly on the quality of the upstream 2D pose detector and fails when 2D keypoints collapse under severe occlusion (Yang et al., 4 Aug 2025). The starshade system only estimates lateral position, not full 3D pose, and regions where the Arago spot is blocked by the central obstruction or off the pupil edge produce larger errors and uncertainties (Chen et al., 2022). DLVS$(x,y)$12-HST-V1 is currently single-target and initially rigid-body only, with no reported baseline pose metrics yet (Velkei et al., 15 Jun 2025). GAP-GDRNet is single-target and purely synthetic, and the paper explicitly states that sim-to-real adaptation remains to be addressed (Zhang et al., 2 Jul 2026). SPNv3 assumes a known target and known 3D keypoints, and its robustness claims are tied to that object-specific setup (Park et al., 2024). Stellar StarPose-like models face fundamental identifiability limits such as north-south flip symmetries, spot number-versus-contrast degeneracy, and the restricted resolution imposed by finite $(x,y)$13 or mesh density (Sagynbayeva et al., 30 Apr 2025, Petralia et al., 2024).

A second common misconception is that these systems are interchangeable because they share the word "pose." The supplied literature shows the opposite. In spacecraft work, pose usually means rigid relative $(x,y)$14. In starshade sensing, it can mean only $(x,y)$15. In stellar-surface inference, it can mean inclination, obliquity, and a statistical map of activity. The unifying idea is therefore methodological rather than ontological: StarPose denotes a class of inference problems in which observations are mapped to latent geometry under strong physical constraints.

This suggests a broader interpretation of the term. StarPose is most coherent when understood as a cross-domain research pattern: combine structured latent geometry, physically informed forward or denoising models, and task-specific uncertainty handling to infer state from incomplete observations. The exact mathematical object being inferred—joints, rigid transforms, pupil-plane offsets, or stellar surfaces—changes by field, but the computational grammar remains strikingly consistent.