EgoTraj-Bench: Towards Robust Trajectory Prediction under Ego-view Noisy Observations

Overview

EgoTraj-Bench is a real-world benchmark for pedestrian trajectory prediction under ego-centric noisy observations. Built upon the TBD dataset, it pairs noisy first-person-view (FPV) derived trajectories with clean bird's-eye-view (BEV) ground truth, enabling robust evaluation under deployment-realistic perception noise.

Data Structure

Level	Folder	Description	Size
L2	L2-processed/	Ready-to-use .npz files for training/evaluation	~44 MB
L1	L1-intermediate/	Core CSV/TXT intermediates: BEV GT, FPV detections/tracks, visibility metadata, robot paths	~302 MB
L0	L0-raw/	Link to TBD raw dataset (~170 GB)	See README

L2: Processed Data

Start here if you want to train or evaluate trajectory prediction models.

L2-processed/
├── EgoTraj-TBD/
│   ├── egotraj_tbd_train.npz
│   ├── egotraj_tbd_val.npz
│   └── egotraj_tbd_test.npz
└── T2FPV-ETH/
    └── t2fpv_{fold}_{split}.npz

Each .npz file contains:

{
    "all_obs":       np.array [N, 8, 7],   # Noisy FPV history
    "all_pred":      np.array [N, 20, 7],  # Clean BEV trajectory
    "num_peds":      np.array [S],
    "seq_start_end": np.array [S, 2],
}
# 7 features = [x, y, orientation, img_x, img_y, valid_mask, agent_id]

L1: Intermediate Data

Use L1-intermediate/ if you want frame/segment-level analysis or to build new data-processing variants.

L1-intermediate/
├── bev_gt/
│   ├── segments/                 # clean BEV pedestrian GT at 2.5 fps
│   └── projected_visibility/     # BEV GT projected into FPV with visibility metadata
├── fpv_detections/
│   ├── segments/                 # segment-level YOLOv8 + BoTSORT FPV tracks
│   └── merged/                   # scene-level merged FPV-derived noisy trajectories
├── robot_paths/                  # ego/robot paths for segment windows
├── scene_splits.json
└── CHECKSUMS.sha256

The released L1 core does not include a standalone matched/ directory. Hungarian matching is materialized in the final L2 tensors. The fpv_detections/merged/ files are FPV detection/tracking outputs projected to BEV/world coordinates, not explicit BEV-FPV assignment records.

L0: Raw Data

Raw TBD data is hosted by the original TBD authors. See L0-raw/README.md.

Quick Start

import numpy as np

data = np.load("L2-processed/EgoTraj-TBD/egotraj_tbd_test.npz")
noisy_history = data["all_obs"][:, :, :2]
clean_past = data["all_pred"][:, :8, :2]
clean_future = data["all_pred"][:, 8:, :2]
valid_mask = data["all_obs"][:, :, 5]

Dataset Details

EgoTraj-TBD

• Source: TBD dataset, 17 recording sessions
• Perception: YOLOv8 detection + BoTSORT tracking on ego-view video
• Sampling: 2.5 fps, 8-frame observation + 12-frame prediction
• L2 Statistics: 36,947 sequences, FPV noisy rate 0.37, history MSE 0.66 m

T2FPV-ETH

• Source: T2FPV simulated ego-centric noise on ETH-UCY
• Folds: eth, hotel, univ, zara1, zara2
• Version: balanced variant (original_bal)

Citation

@inproceedings{liu2025egotraj,
  title={EgoTraj-Bench: Towards Robust Trajectory Prediction under Ego-view Noisy Observations},
  author={Liu, Jiayi and Zhou, Jiaming and Ye, Ke and Lin, Kun-Yu and Wang, Allan and Liang, Junwei},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2025}
}

License

This dataset is released under CC BY-NC 4.0. The underlying TBD raw data is subject to its own license; please refer to the TBD dataset page.