Interactive Pose Priors

• Interactive pose priors are statistical, generative, or learned constraints that condition pose estimation on rich temporal, spatial, semantic, or multimodal context.
• They integrate cues like action class and interaction signals to infer, predict, and refine physically plausible poses in dynamic or ambiguous scenarios.
• State-of-the-art methods employ diffusion models, transformer architectures, and normalizing flows to enhance robustness in multi-object, multi-agent, and action-conditioned contexts.

Interactive pose priors are statistical, generative, or learned constraints that encode context-dependent knowledge about body configuration, interaction, or motion, facilitating robust and physically plausible pose estimation, motion synthesis, and action understanding. Unlike static priors applied to individual configurations, interactive pose priors incorporate temporal, spatial, semantic, or multimodal contextual information—such as action class, interaction cues, or multimodal embeddings—enabling the system to infer, predict, or refine pose estimates in dynamic or ambiguous scenarios.

1. Core Concepts and Formal Definitions

Interactive pose priors generalize beyond traditional marginal priors by conditioning the pose distribution on rich context. In graphical models, context might include high-level activity labels (Iqbal et al., 2016), action hypotheses inferred from current or past predictions (Iqbal et al., 2016), language cues (Subramanian et al., 6 May 2024), or spatial proximity in human–human interaction (Liu et al., 16 Oct 2025). The formulation is typically as follows:

$$p(X \mid C, I) \propto \prod_{j \in J} \phi_j(x_j \mid C, I) \prod_{(j, p) \in E} \psi_{jp}(x_j, x_p \mid C)$$

where $X$ is the pose configuration, $C$ is the contextual prior (e.g., action, language, or contact), $\phi_j$ are appearance unaries, and $\psi_{jp}$ are spatial/binary terms. Modern approaches often replace or augment this with diffusion models (Liu et al., 16 Oct 2025, Ta et al., 18 Oct 2024, Lu et al., 2023, Ci et al., 2022), transformer-based architectures (Zhang et al., 2023, Xu et al., 25 Feb 2025), or normalizing flows (Heker et al., 16 Jul 2025), each enabling conditioning on rich or multi-modal context.

Interactive pose priors may be explicit—such as action semantic class influencing pose-likelihood (Iqbal et al., 2016)—or implicit, as in score-based or generative models where context reshapes the high-density manifold the model learns to sample from or denoise towards (Ci et al., 2022, Lu et al., 2023, Ta et al., 18 Oct 2024, Liu et al., 16 Oct 2025).

2. Action-Conditioned and Contextual Pose Priors

A prototypical example is the Action-Conditioned Pictorial Structure (ACPS) model (Iqbal et al., 2016), which integrates an action prior $p_A(a)$ into both the unary and binary terms of the articulated body configuration model. The system starts with a uniform action prior, estimates pose sequences, infers actions from these sequences using a bag-of-words pose descriptor, and then re-injects the derived action prior back into the pose estimation pipeline:

• The conditional unary:

$$\phi_j(x_j \mid p_A, I) = \sum_{a \in A} p_A(a) \phi_j(x_j \mid a, I)$$

• The binary potentials are conditioned on the most probable action class to reflect activity-specific kinematic constraints.

Furthermore, the ACPS model introduces appearance sharing among action classes through learned weights $\gamma_a(a')$, improving robustness to misclassified or ambiguous actions by blending appearance models across semantically similar activities.

Similar approaches are observed in action-driven pose estimation and physics-based simulation, where motion priors are constructed for specific parts or activities, then reassembled or conditioned according to the interaction context (Bae et al., 2023).

3. Generative and Diffusion-based Interactive Priors

Recent advances employ generative diffusion models or normalizing flows to construct expressive, context-sensitive pose priors. In these paradigms, a noisy pose is progressively denoised or transformed toward a distribution shaped jointly by unimodal pose statistics and context-specific signals:

• DPoser (Lu et al., 2023) and MOPED (Ta et al., 18 Oct 2024) employ diffusion processes for robust human pose priors, modeling $p_\text{pose}(x)$ with both unconditional and conditional (multi-modal) inference, where context $c$ can be text, image, or partial keypoints.
• In DPoser, the regularizer term

$$L_\text{DPoser} = w_t \| x_0 - \text{sg}[\hat{x}_0(t)] \|^2$$

incorporates a one-step denoising prediction, enforcing proximity to the learned pose manifold at each optimization step.

These generative models can be conditioned on textual input, image embeddings (using CLIP features), partial observations (for completion/denoising), or structured prior knowledge (activity labels, object category) (Ta et al., 18 Oct 2024, Lu et al., 2023, Liu et al., 16 Oct 2025), significantly improving both diversity and plausibility of synthesized or completed poses.

4. Interactive Priors for Multi-Object and Multi-Agent Scenarios

Interactive pose priors extend naturally to hand–object (Zhu et al., 2023), multi-agent, or human–human interaction settings (Liu et al., 16 Oct 2025). In these domains, priors must capture contact, proximity, and dynamic dependencies:

• ContactArt (Zhu et al., 2023) learns both an articulation prior for object part arrangement (via GAN) and a contact prior (via diffusion modeling of contact maps), enabling robust hand–object pose estimation and transferability to real-world domains.
• Ponimator (Liu et al., 16 Oct 2025) models the joint distribution of proximal interactive poses and their temporal context using two diffusion models: one for generating spatially plausible interactive poses from a single actor, pose, or text description; one for animating realistic multi-person motion sequences from an anchor interactive pose.

In language-conditioned frameworks, ProsePose (Subramanian et al., 6 May 2024) operationalizes interactive priors by extracting contact constraints from large multimodal LLMs and translating them into differentiable losses for pose graph or mesh optimization, facilitating zero-shot pose completion and refinement in interaction-rich scenes.

5. Adaptive and Unsupervised Learning of Interactive Priors

Several works address the challenge of learning interactive pose priors without strong supervision. The Pose Prior Learner (PPL) (Wang et al., 4 Oct 2024) uses compositional hierarchical memory and unsupervised training to extract prototypical poses for any object category, relying only on image reconstruction losses:

• The memory module stores sub-structures of prototypical poses.
• Iterative inference regresses initial pose proposals back toward these learned prototypes, making the pose estimation robust to occlusions and uncertainty.

In animal and cross-category pose settings, transformers equipped with keypoint clustering and body part prompts provide adaptive, instance-aware prior information that is dynamically fused for each prediction (Xu et al., 25 Feb 2025).

6. Practical Applications and Empirical Results

Interactive pose priors have been empirically validated across domains:

Method	Core Context Integration	Application Domains	Reported Impact
ACPS (Iqbal et al., 2016)	Action priors, appearance sharing	Monocular human pose estimation	+~4% APK on sub-J-HMDB, +6% on Penn-Action
GFPose (Ci et al., 2022)	Score-based diffusion, conditioning	3D human pose estimation, completion	~20% improvement in minMPJPE (Human3.6M)
MOPED (Ta et al., 18 Oct 2024)	Multi-modal conditioning	Pose estimation, denoising, completion	Lower PA-MPJPE and higher diversity over VPoser, DPoser
ContactArt (Zhu et al., 2023)	Interaction priors (GAN, diffusion)	Hand–object pose, transfer to real data	Higher IoU, 5°/5cm, lower error on hand/object datasets
Ponimator (Liu et al., 16 Oct 2025)	Spatial/temporal priors, diffusion	Human–human animation, image/text 2 motion	Increased contact, reduced penetration, better realism
ProsePose (Subramanian et al., 6 May 2024)	Language-model contact constraints	Multi-person, self-contact pose estimation	Increased correct contact, narrows gap to supervised ALT

These frameworks support tasks including robust pose estimation under occlusion, interactive animation, pose completion, virtual try-on (Liang et al., 22 Aug 2024), robotics (Liu et al., 14 Feb 2025), and hand motion tracking (Duran et al., 2023).

7. Future Directions and Open Challenges

The field is trending toward richer, more flexible interactive priors via:

• End-to-end joint learning of priors and predictors using large-scale multimodal, multi-agent datasets.
• More efficient sampling or inference—accelerated denoising steps, approximate diffusion, or real-time transformer pipelines (Ta et al., 18 Oct 2024, Liu et al., 16 Oct 2025).
• Dynamic priors that adapt to streaming video, user input, or novel scene context without expensive retraining (Adjel, 21 Jul 2025).
• Integration of symbolic, language, or programmatic constraints for human–robot collaboration, semantic activity prediction, or self-supervised lifelong learning (Subramanian et al., 6 May 2024, Wang et al., 4 Oct 2024).
• Interactive systems where user feedback or downstream performance can iteratively refine the pose prior for improved robustness and personalization.

A plausible implication is that future interactive pose estimation pipelines will seamlessly blend discriminative predictions, generative priors, and external semantic knowledge to enable physically plausible, context-aware, and adaptable solutions across an even broader range of vision, graphics, and robotics applications.