Research map / Label-efficient and robust learning / Self-supervised and contrastive
Enhancing Trajectory Prediction through Self-Supervised Waypoint Distortion Prediction
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 SSWNP, a plug-in self-supervised method for trajectory prediction that enforces consistency between clean and noise-augmented waypoint views and adds an auxiliary noise-prediction task, improving GroupNet, AutoBot, Graph-TERN and SSAGCN on NBA, TrajNet++ and ETH-UCY.
Problem
Trajectory prediction models tend to produce overly simplified predictions because they rely on a narrow interpretation of the data manifold, and they degrade on noisy inputs.
Approach
Create a clean view and a Gaussian noise-augmented view of observed waypoints, train the predictor so both views yield predictions close to ground truth (spatial consistency), and add an auxiliary MLP that predicts the injected noise from features as a self-supervised pretext task.
Key results
- On NBA, GroupNet+SSWNP reaches 0.90 ADE / 1.14 FDE at 4.0s versus 1.13/1.69 for GroupNet (22.7%/38.9% relative gain).
- On TrajNet++, AutoBot+SSWNP reaches 0.091 minADE / 0.162 minFDE versus 0.128/0.234 for AutoBot (33.8%/36.4% gain).
- On ETH-UCY, Graph-TERN+SSWNP averages 0.22/0.33 versus 0.24/0.38, and SSAGCN+SSWNP averages 0.11/0.19 versus 0.13/0.24.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SSWNP to GroupNet | GroupNet · NBA | ADE/FDE (4.0s): 1.13/1.69 -> 0.90/1.14; 22.7/38.9 | Helped |
| Add SSWNP to AutoBot | AutoBot · TrajNet++ | Scene-level Min ADE/Min FDE: 0.128/0.234 -> 0.091/0.162; 33.8/36.4 | Helped |
| Add SSWNP to Graph-TERN | Graph-TERN · ETH-UCY AVG | ADE/FDE: 8.60/14.00 | Helped |
| Add SSWNP to SSAGCN | SSAGCN · ETH-UCY AVG | ADE/FDE: 16.60/23.20 | Helped |
| Add SSWNP to AutoBot | AutoBot · TrajNet++, noisy test environment | ADE/FDE: 0.301/0.469 -> 0.134/0.195; 76.8/82.5 | Helped |
| Add SSWNP to GroupNet | GroupNet · NBA, noisy test environment | ADE/FDE: 1.784/1.771 -> 0.95/1.23; 61.0/36.1 | Helped |
| Add spatial consistency to GroupNet (ablation) | GroupNet · NBA | ADE/FDE (4.0s): 1.13/1.69 -> 1.018/1.362 | Helped |
| Add noise prediction to GroupNet with spatial consistency (ablation) | GroupNet + spatial consistency · NBA | ADE/FDE (4.0s): 1.018/1.362 -> 0.903/1.147 | Helped |
| Set noise factor to 0.05 (ablation) | GroupNet + SSWNP · NBA, validation | ADE/FDE (4.0s): 0.896/1.130 | Helped |
| Set noise factor to 0 instead of 0.05 (ablation) | GroupNet + SSWNP · NBA, validation | ADE/FDE (4.0s): 0.896/1.130 -> 1.13/1.69 | Hurt |
Surprising findings
- Spatial consistency alone (B+SC) gives a clear gain, but the paper reports that the auxiliary noise prediction task adds further improvement on NBA.
- Reported gains under noisy test inputs (up to 76.8% ADE on TrajNet++) are much larger than on clean inputs, though the noise protocol is only briefly described.
Limitations
- Noise factor and lambda must be tuned per dataset, with values varying widely across datasets (0.001 to 0.1).
- Gains on ETH-UCY are uneven; for SSAGCN, ADE on ETH is unchanged and the reported averages are based on author-selected baselines.
- Tables mix results reported as minADE/minFDE and ADE/FDE, and some table entries are inconsistent, weakening comparability.
Experimental setup
Datasets: NBA SportVU TrajNet++ ETH-UCY
Benchmarks: NBA SportVU TrajNet++ ETH-UCY
Metrics: ADE FDE minADE minFDE
Baselines: GroupNet AutoBot Graph-TERN SSAGCN DynGroupNet MERA DISTL Stimulus TDGCN SIM Trajectron++ PecNet
Builds on: GroupNet AutoBot Graph-TERN SSAGCN
Closest papers by meaning
- STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.87)
- T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory (CVPR 2024 · similarity 0.86)
- HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention (CVPR 2024 · similarity 0.84)
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.84)
- SEPT: Towards Efficient Scene Representation Learning for Motion Prediction (ICLR 2024 · similarity 0.84)
- Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM (CVPR 2025 · similarity 0.83)
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.