ECCV 2026 Spotlight · Malmö, Sweden

SPARC Single-Pass Scaling for Motion Forecasting with Conformal Bayesian Last Layers

Sakif Hossain* · Julian Teusch* · Jörg P. Müller * Equal contribution

A fast deterministic motion forecaster becomes a structured, calibrated predictive density without Monte Carlo sampling or an ensemble loop.

Paper on arXiv Code on GitHub Visual notebook Presentations Supplement

The problem

Point forecasts do not say when they should be trusted.

Human motion forecasting is a core component for planning around people. SPARC retains the speed of a deterministic backbone while adding an analytic epistemic signal, structured graph-temporal covariance and post-hoc split conformal calibration.

1.00NLL mean rank

2.69MPJPE + NLL mean rank

1.79×MPJPE in the highest-κ decile

1forward pass at inference

Method

Structured uncertainty, calibrated in one pass.

Bayesian leverage

A conjugate Bayesian last layer yields κₜ(x), an analytic measure of how strongly a test input is supported by the learned feature representation.

Preserved structure

κₜ scales a graph-temporal covariance. Uncertainty expands where support is weak while correlations across joints and future steps remain intact.

Conformal calibration

Held-out residual quantiles turn predictive scales into 95% marginal tubes with finite-sample validity under exchangeability.

Animated explanations

Watch uncertainty take shape.

The technical deck contains five animated figures. They remain live here, including covariance scaling, calibration and low-versus-high κ examples.

Animated comparison of calibrated intervals for low and high kappa — Higher κ produces visibly wider calibrated intervals.

Animation comparing diagonal and structured uncertainty tubes — Scale epistemics while preserving trajectory structure.

Animation of temporal covariance inflation — Inflate temporal covariance where epistemic leverage is high.

Animation of split conformal calibration — Convert held-out residuals into calibrated 95% tubes.

Animation comparing efficient and overly wide calibrated intervals — Target coverage without needlessly wide intervals.

Technical presentation

Explore both SPARC presentations.

Switch between the extended technical deck and the updated 5-minute spotlight. Animated figures remain live inside the extended presentation.

1 / 42

ECCV 2026

The project at a glance.

The conference poster condenses the motivation, six-stage pipeline, main benchmark results and deployment scope into a single visual overview.

Resources

Paper, code and presentations.

SPARC-UQ on GitHubReusable core for point estimators · Apache 2.0 · not the full benchmark implementation Python package · v0.1.0Installable GitHub wheel and source archive · alpha release · PyPI pending Executed notebook · six visualizationsSynthetic forecasts, coverage, covariance and feature-leverage diagnostics Plotting and diagnostic APIReusable Matplotlib helpers · signed scale contributions, not causal explanations Paper on arXivarXiv:2608.20802 Supplementary materialExtended methods and results Conference posterECCV 2026 · 1400 × 1000 mm Extended technical deck42 slides · full technical deck ECCV spotlight deck5-minute talk · 11 slides

Citation

Accepted at ECCV 2026.

Sakif Hossain*, Julian Teusch*, and Jörg P. Müller. SPARC: Single-Pass Scaling for Motion Forecasting with Conformal Bayesian Last Layers. European Conference on Computer Vision (ECCV), 2026. *Equal contribution.