Pose Structure Score (PSS) Metric

• Pose Structure Score (PSS) is a scale-invariant metric that evaluates 3D human pose plausibility by checking if predicted and ground-truth poses belong to the same canonical cluster.
• It employs a workflow of normalization, k-means clustering, and binary assignment to address limitations of traditional joint-wise error metrics.
• Beyond computer vision, PSS has been adapted to bioinformatics and drug discovery, linking geometric correctness with biological or physicochemical relevance.

The Pose Structure Score (PSS) is a scale-invariant, structure-aware performance metric introduced to evaluate the structural plausibility of predicted 3D human poses relative to ground truth reference poses. PSS distinguishes itself from joint-wise distance metrics by assessing whether predicted and ground-truth poses fall within the same canonical cluster in a learned pose vocabulary. The metric provides insight into the global geometric configuration of a pose, addressing limitations of traditional evaluative measures that may miss semantically significant but spatially small deviations. PSS has also been adapted as a generic concept of pose quality in structural bioinformatics and drug discovery. Its binary, cluster-based formalism serves as a bridge between strict geometric invariance and biologically or physically meaningful correctness in pose prediction tasks (Kocabas et al., 2019, Gniewek et al., 2021).

1. Motivation and Rationale

Traditional pose estimation benchmarks such as Mean Per-Joint Position Error (MPJPE) and Percentage of Correct Keypoints (PCK) measure per-joint distances independently. These metrics can equate poses with similar absolute errors but very different structural semantics. For tasks where overall limb configuration and the plausibility of articulation are intrinsic—such as action recognition or robotics—correct structure supersedes coordinate-level proximity. PSS operationalizes this distinction by determining if a predicted pose shares the same pose “type” (cluster) as its ground truth under a learned, scale-invariant vocabulary of poses, irrespective of global scale or translational offsets (Kocabas et al., 2019).

2. Formal Definition and Computation

Let $\mathbf{q}_i \in \mathbb{R}^{3J}$ for $i=1,\dots,n$ denote a large set of $n$ reference ground-truth poses, where each pose is concatenated as a $3J$-dimensional vector corresponding to joint coordinates. The construction of PSS proceeds via the following steps:

1. Normalization: Each pose is centralized (typically made root-relative) and scaled to unit length:

$$\hat{\mathbf{q}}_i = \frac{\mathbf{q}_i}{\|\mathbf{q}_i\|_2}$$

1. Clustering: $k$-means clustering is performed over $\{\hat{\mathbf{q}}_i\}$, yielding $k$ cluster centers $\{\boldsymbol\mu_1, \dots, \boldsymbol\mu_k\}$.
2. Assignment Function: For an arbitrary pose $\mathbf{x}$, define the cluster assignment:

$$C(\mathbf{x}) = \arg\min_{1 \le m \le k} \left\| \frac{\mathbf{x}}{\|\mathbf{x}\|_2} - \boldsymbol\mu_m \right\|_2^2$$

1. Score Computation: For a predicted pose $\mathbf{p}$ and its corresponding ground truth $\mathbf{q}$,

$$\mathrm{PSS}(\mathbf{p}, \mathbf{q}) = \delta(C(\mathbf{p}), C(\mathbf{q}))$$

where

$$\delta(i, j) = \begin{cases} 1 & \text{if } i = j \ 0 & \text{otherwise} \end{cases}$$

The mean-PSS (mPSS) over a test set $\{(\mathbf{p}_t, \mathbf{q}_t)\}_{t=1}^T$ is computed as:

$$\mathrm{mPSS} = \frac{1}{T}\sum_{t=1}^{T} \mathrm{PSS}(\mathbf{p}_t, \mathbf{q}_t)$$

(Kocabas et al., 2019).

3. Preprocessing and Algorithmic Workflow

The PSS pipeline is split into offline and online phases. Offline, reference poses are normalized and clustered to generate the canonical pose vocabulary. This is performed once per dataset. Online, each predicted/ground-truth pose pair is normalized and assigned to its nearest cluster center. The resulting PSS quantifies their structural concordance as a binary value.

Offline Computation

Input: reference ground-truth poses {q_i} for i=1…n, number of clusters k 1. For each q_i, compute normalized pose: q̂_i ← q_i / ||q_i||_2 2. Run k-means on {q̂_i} to obtain centers {μ_1,…,μ_k} Output: cluster centers {μ_m}

Online Computation

function PSS(p, q, {μ_m}): # 1. Normalize both poses p̂ ← p / ||p||_2 q̂ ← q / ||q||_2 # 2. Find nearest center indices c_p ← argmin_{m=1..k} || p̂ – μ_m ||_2^2 c_q ← argmin_{m=1..k} || q̂ – μ_m ||_2^2 # 3. Score structural agreement if c_p == c_q: return 1 else: return 0 end

(Kocabas et al., 2019)

4. Properties and Theoretical Characteristics

• Binary Per-Sample Output: Each pose pair returns either 0 or 1.
• Range and Interpretation: mPSS $\in [0,1]$, commonly reported as a percentage; higher is better.
• Scale and Translation Invariance: All poses are normalized and centralized, obviating the influence of global scale, translation, and sign reversal.
• Sensitivity to Structural Deviations: Joint angle errors, abnormal limb configurations, or semantic misalignments typically induce cluster assignment changes, resulting in a lower PSS even if Euclidean errors remain small.
• Stability Under Clustering: Extensive k-means reruns establish high cluster stability (mean IOU = 0.78; mPSS variance $< \pm$0.1%), evidencing the robustness of the PSS metric to initializations (Kocabas et al., 2019).

5. Experimental Validation and Comparative Performance

Empirical evaluation on benchmark datasets underpins the structural sensitivity of PSS:

Setting	mPSS@50	mPSS@100
H36M, Full Supervision	84.44%	78.67%
H36M, EpipolarPose (Self-Sup.)	73.09%	64.03%
H36M, EpipolarPose + Refinement	80.42%	75.41%
H36M, 2D GT Triangulation	83.9%	—
MPI-INF-3DHP, Full Supervision	87.15%	82.21%
MPI-INF-3DHP, SS subj-1 only	75.64%	73.15%
MPI-INF-3DHP, SS no labels	70.94%	67.58%

Results indicate that PSS decreases as 2D detector quality declines or as supervision is limited. Notably, models yielding similar MPJPE can diverge meaningfully in PSS, demonstrating cases where MPJPE underestimates structural implausibility. For instance, even with perfect 2D keypoints and camera extrinsics, near-maximal PSS is obtained (≈99%), whereas degraded input can drop mPSS by up to 10 percentage points (Kocabas et al., 2019).

6. Extension to Pose Quality in Structural Bioinformatics

Analogous concepts are adapted to structure-based virtual screening (vHTS) domains. The “pose-quality score” in (Gniewek et al., 2021) is derived from AtomNet PoseRanker and predicts pose plausibility as an explicit regression target in a multi-task GCN. This score is used both as an auxiliary learning target and as a conditioning variable for bioactivity prediction, directly influencing the model’s pose-sensitivity. Incorporating poor poses of known actives as negatives and explicitly conditioning activity prediction on pose quality results in pronounced sensitivity to geometric correctness, as measured by drops in predicted activity (Δ_drop) between good, poor, and implausible poses—metrics analogous in spirit to PSS’s structural focus. Benchmarks introduced in this context—such as the Picasso Problem on hZAP70—quantify a model’s ability to penalize inaccurate binding geometries, emphasizing the broad applicability of pose-structure-based evaluation methodologies (Gniewek et al., 2021).

7. Significance and Implications

Pose Structure Score provides an indispensable complement to traditional error-based metrics in both computer vision and computational biology. By penalizing semantically meaningful pose errors, PSS better aligns evaluation with task-specific demands where articulation rather than marginal joint error is critical. Its simplicity, scalability, and demonstrated stability recommend it as a universal diagnostic for both supervised and self-/weakly-supervised learning settings. Furthermore, the generic concept of pose-structure-aware scoring underpins advances in virtual screening and protein-ligand interaction prediction, where explicit pose quality conditioning improves model robustness and interpretability. A plausible implication is the proliferation of structurally-informed evaluation as a standard for pose-driven tasks across modalities (Kocabas et al., 2019, Gniewek et al., 2021).