Research map / Detection and segmentation / Tracking
GigaTraj: Predicting Long-term Trajectories of Hundreds of Pedestrians in Gigapixel Complex Scenes
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.
Introduces GigaTraj, a gigapixel, wide-FoV pedestrian trajectory dataset with multimodal annotations, and benchmarks existing trajectory predictors on minute-level long-term prediction.
Problem
Existing pedestrian trajectory datasets lack wide field of view combined with high resolution, so minute-level long-term trajectory prediction in large crowded scenes cannot be studied.
Approach
A gigapixel camera-array dataset of 16 scenes (~4x10^4 m^2) extended from PANDA with homography-based world coordinates, group/interaction graphs, scene semantics, and a train/test split including unseen scenes, plus an empirical benchmark of trajectory predictors at 30s/10s horizons.
Key results
- MemoNet achieves overall minADE20/minFDE20 of 2.04/2.87 at 30s horizon, and 0.89/1.00 at 10s horizon on GigaTraj.
- Trajectron++ reaches 1.72/2.85 overall at 30s but degrades to 22.40/53.67 on Scene 12, showing poor generalization to unseen scenes.
- In PECNet ablation, adding graph information reduces minADE20/minFDE20 from 3.85/6.85 to 2.18/4.71.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add graph information (ablation) | PECNet · GigaTraj test | min ADE20: 3.85 -> 2.18 | Helped |
| add graph information (ablation) | PECNet · GigaTraj test | min FDE20: 6.85 -> 4.71 | Helped |
| add scene semantics as input (ablation) | PECNet without graph information · GigaTraj test | min ADE20: 3.85 -> 3.72 | Helped |
| add scene semantics as input (ablation) | PECNet without graph information · GigaTraj test | min FDE20: 6.85 -> 6.64 | Helped |
| increase prediction horizon from 10s to 30s | trajectory prediction baselines · GigaTraj test | prediction error | Hurt |
| use simple social pooling in Social-LSTM and SGAN rather than Vanilla-LSTM | GigaTraj trajectory prediction · GigaTraj test, overall (T=30s) | prediction error | Hurt |
Surprising findings
- Social-LSTM and SGAN, which model pedestrian interactions, yield larger errors than vanilla LSTM that ignores relationships.
- Concatenating scene semantics to PECNet without graph information did not improve results.
Limitations
- Scenes 15 and 16 were not fully annotated at submission, so overall results use partial test sets.
- Baselines were re-implemented from descriptions rather than official code, except PECNet and MemoNet.
- Methods use ground-truth coordinates and BEV maps rather than detection/tracking from raw images.
Experimental setup
Datasets: GigaTraj PANDA
Benchmarks: GigaTraj
Metrics: ADE FDE minADE20 minFDE20
Baselines: Vanilla-LSTM Social-LSTM SGAN PECNet Trajectron++ MemoNet
Builds on: PANDA Social-LSTM Social GAN PECNet Trajectron++ MemoNet
Closest papers by meaning
- Towards Predicting Any Human Trajectory In Context (NeurIPS 2025 · similarity 0.86)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · similarity 0.86)
- MS-TIP: Imputation Aware Pedestrian Trajectory Prediction (ICML 2024 · similarity 0.86)
- A Unified Environmental Network for Pedestrian Trajectory Prediction (AAAI 2024 · similarity 0.85)
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.85)
- MotionMap: Representing Multimodality in Human Pose Forecasting (CVPR 2025 · similarity 0.85)
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.