Research map / Probabilistic and scientific modeling / Molecules and proteins
SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SocialCVAE, a hybrid pedestrian trajectory predictor whose CVAE is conditioned on an energy-based interaction map that encodes anticipated collision repulsion, improving ADE/FDE on ETH-UCY and SDD.
Problem
Pedestrian trajectory prediction: explainable empirical (energy/force-based) models are deterministic and inaccurate, while deep models are black boxes; hybrid models suffer from deterministic steering behavior. The paper aims to model socially reasonable behavioral uncertainty with interpretability
Approach
Use an energy-based interaction model to build a local interaction energy map (anticipated repulsion/occupancy from neighbors, static and dynamic obstacles) and use it as the condition of a CVAE that learns residuals over a coarse attention-based velocity prediction, so the sampled randomness is socially explainable.
Key results
- On SDD, SocialCVAE reaches ADE20/FDE20 of 1.48/1.06 pixels versus 1.78/3.44 for NSP-SFM, a reported 16.85% ADE and 69.18% FDE improvement.
- On ETH-UCY, average ADE20/FDE20 of 0.03/0.02 meters, with a reported 66.67% FDE improvement over the prior SOTA.
- Ablation on SDD: removing the interaction-conditioned CVAE raises ADE/FDE from 1.48/1.06 to 1.76/1.57; replacing the attention coarse model with goal attraction gives 1.56/2.71.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SocialCVAE instead of NSP-SFM | pedestrian trajectory prediction · ETH-UCY AVG | FDE: 0.06 -> 0.02 | Helped |
| use SocialCVAE instead of NSP-SFM | pedestrian trajectory prediction · SDD | ADE: 1.78 -> 1.48 | Helped |
| use SocialCVAE instead of NSP-SFM | pedestrian trajectory prediction · SDD | FDE: 3.44 -> 1.06 | Helped |
| add the attention model to the coarse prediction model (ablation) | SocialCVAE · SDD | ADE: 8.64 -> 1.48 | Helped |
| add the attention model to the coarse prediction model (ablation) | SocialCVAE · SDD | FDE: 13.72 -> 1.06 | Helped |
| condition the CVAE on the interaction energy map as well as past trajectories (ablation) | SocialCVAE · SDD | ADE: 1.76 -> 1.48 | Helped |
| condition the CVAE on the interaction energy map as well as past trajectories (ablation) | SocialCVAE · SDD | FDE: 1.57 -> 1.06 | Helped |
| use the proposed coarse motion prediction model instead of the NSP-SFM goal-attraction model (ablation) | SocialCVAE · SDD | ADE: 1.56 -> 1.48 | Helped |
| use the proposed coarse motion prediction model instead of the NSP-SFM goal-attraction model (ablation) | SocialCVAE · SDD | FDE: 2.71 -> 1.06 | Helped |
Surprising findings
- Reported inference is computationally expensive: 336.62 s for the SDD test set with 20 samples per step, dominated by per-pedestrian CVAE and energy-map computation.
- FDE is sometimes lower than ADE on several test cases, which the authors attribute to the recursive prediction scheme.
Limitations
- Computationally inefficient because energy maps and CVAE sampling are computed sequentially per pedestrian at each time step.
- Interaction modeling restricted to collision-related energy; other social factors (affective states, traffic signals) are not modeled.
Experimental setup
Datasets: ETH-UCY Stanford Drone Dataset (SDD)
Benchmarks: ETH-UCY SDD
Metrics: ADE20 FDE20
Baselines: S-GAN Sophie Trajectron++ PECNet YNET Social-VAE CAGN SIT MUSE-VAE MSRL S-CSR NSP-SFM
Builds on: NSP-SFM Social-VAE CSR PECNet
Closest papers by meaning
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.90)
- SocialCircle: Learning the Angle-based Social Interaction Representation for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.89)
- SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction (CVPR 2025 · similarity 0.89)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · similarity 0.87)
- Higher-order Relational Reasoning for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.86)
- Towards Predicting Any Human Trajectory In Context (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.