Research map / Label-efficient and robust learning / Recognition and long-tail
Adapting to Observation Length of Trajectory Prediction via Contrastive Learning
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 CLLS, a contrastive learning method using random length masking, rotation and scaling augmentations to learn length-invariant trajectory features, plus a lightweight GRU-based RNLS model, for human trajectory prediction under varying observation lengths.
Problem
Human trajectory prediction models degrade when observation lengths are limited or vary (observation length shift), and existing fixes add architectural complexity.
Approach
Adds a contrastive learning loss during training only (SimCLR-style, with a nonlinear projector) over positive pairs made of length-masked, rotated and scaled versions of the same trajectory, so existing encoders learn length-invariant features with no extra inference cost; combined with a lightweight MLP+GRU network (RNLS) with social cross-attention.
Key results
- On SDD with 8-frame observations, RNLS & CLLS reaches 7.2/11.0 ADE/FDE, versus 7.8/10.1 for LMTraj-SUP (lower ADE than the best prior, but higher FDE).
- With 2-frame observations, CLLS improves Trajectron++ on ETH-UCY average ADE/FDE from 0.29/0.56 (BCDiff) to 0.27/0.46.
- On NBA, RNLS & CLLS reports 0.65/1.20 for rebounding and 0.61/1.09 for scoring, improving over the second-best by 9.7%/12.4% and 4.7%/6.8%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CLLS to Trajectron++ | Trajectron++ · ETH-UCY, observation length 2 | ADE/FDE: 0.29/0.53 -> 0.27/0.46 | Helped |
| Apply CLLS to Trajectron++ | Trajectron++ · SDD, observation length 2 | ADE/FDE: 11.56/19.32 -> 10.05/17.42 | Helped |
| Use RNLS & CLLS instead of SocialVAE | ETH-UCY, full 8-frame observation · ETH-UCY | ADE/FDE: 0.21/0.33 -> 0.20/0.32 | Helped |
| Use RNLS & CLLS instead of LMTraj-SUP | SDD, full 8-frame observation · SDD | ADE/FDE: 7.8/10.1 -> 7.2/11.0 | Mixed |
| Use RNLS & CLLS instead of SocialVAE | NBA Rebounding · NBA Rebounding, full 8-frame observation | ADE/FDE: 0.72/1.37 -> 0.65/1.20 | Helped |
| Use RNLS & CLLS instead of SocialVAE | NBA Scoring · NBA Scoring, full 8-frame observation | ADE/FDE: 0.64/1.17 -> 0.61/1.09 | Helped |
| Add an MLP projector for contrastive loss (ablation) | RNLS & CLLS · SDD, unobservable length 6 | ADE/FDE: 8.23/12.94 -> 7.97/12.54 | Helped |
| Add random length masking (ablation) | RNLS & CLLS · SDD, unobservable length 6 | ADE/FDE: 44.47/75.38 -> 8.05/12.08 | Helped |
| Add scaling augmentation (ablation) | RNLS & CLLS · SDD, unobservable length 6 | ADE/FDE: 8.23/13.03 -> 8.05/12.08 | Helped |
| Add rotation augmentation (ablation) | RNLS & CLLS · SDD, unobservable length 6 | ADE/FDE: 8.16/12.90 -> 8.05/12.08 | Helped |
| Set contrastive-loss weight ω to 10 instead of 0.1 (ablation) | RNLS & CLLS · SDD, unobservable length 6 | ADE/FDE: 7.97/12.54 -> 8.04/12.73 | Mixed |
| Set contrastive-loss weight ω to 1 instead of 0.1 (ablation) | RNLS & CLLS · SDD, unobservable length 6 | ADE/FDE: 7.97/12.54 -> 8.05/12.08 | Mixed |
Surprising findings
- The contrastive method performed poorly on PCCSNet and on ZARA1 with Trajectron++, which the authors attribute to memory-bank interference and rotation augmentation disrupting motion distribution.
- Without length masking, performance at longer unobserved lengths degrades sharply (e.g., 44.47 ADE at Tunobs=6 on SDD).
Limitations
- Contrastive invariance may sacrifice fine-grained features and hurt base performance.
- Length-masking, scaling and rotation augmentations are ineffective for map-based data.
- Underperforms on PCCSNet and on ZARA1 for Trajectron++.
Experimental setup
Datasets: ETH-UCY SDD NBA
Benchmarks: ETH-UCY SDD NBA
Metrics: ADE FDE
Baselines: IT DTO MOE BCDiff Social GAN Social-STGCNN PECNet Trajectron++ AgentFormer SocialVAE EqMotion LED
Builds on: SimCLR SocialVAE BCDiff Trajectron++ winner-takes-all
Closest papers by meaning
- Adapting to Length Shift: FlexiLength Network for Trajectory Prediction (CVPR 2024 · similarity 0.88)
- Contrast encodes inductive bias: separating slow noise from dynamics in predictive representation learning (NeurIPS 2026 · similarity 0.85)
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (NeurIPS 2024 · similarity 0.84)
- TrajCLIP: Pedestrian trajectory prediction method using contrastive learning and idempotent networks (NeurIPS 2024 · similarity 0.84)
- Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making (ICML 2024 · similarity 0.84)
- SingularTrajectory: Universal Trajectory Predictor Using Diffusion Model (CVPR 2024 · similarity 0.84)
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.