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

Overview

EgoTraj-Bench is the first 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 of trajectory prediction models under deployment-realistic conditions.

Data Structure

This repository provides data at three levels:

Level	Folder	Description	Size
L2	L2-processed/	Ready-to-use .npz files for training/evaluation	~44 MB
L1	L1-intermediate/	Frame-level FPV detections + BEV GT CSVs + matching results	Coming soon
L0	L0-raw/	Link to TBD raw dataset (~170 GB)	See README

L2: Processed Data (Start Here)

L2-processed/
├── EgoTraj-TBD/                       # Real-world ego-centric noise (17 recording sessions)
│   ├── egotraj_tbd_train.npz
│   ├── egotraj_tbd_val.npz
│   └── egotraj_tbd_test.npz
└── T2FPV-ETH/                         # Simulated ego-centric noise (5 folds)
    └── t2fpv_{fold}_{split}.npz       # folds: eth, hotel, univ, zara1, zara2

Each .npz file contains:

{
    "all_obs":       np.array [N, 8, 7],   # Noisy FPV history (8 past frames)
    "all_pred":      np.array [N, 20, 7],  # Clean BEV trajectory (8 past + 12 future)
    "num_peds":      np.array [S],          # Number of agents per scene
    "seq_start_end": np.array [S, 2],       # Scene boundaries in agent dimension
}
# 7 features = [x, y, orientation, img_x, img_y, valid_mask, agent_id]
# x, y: world coordinates (meters)
# valid_mask: 1 = FPV observation valid, 0 = occluded/missing

Quick Start

import numpy as np

data = np.load("L2-processed/EgoTraj-TBD/egotraj_tbd_test.npz")
noisy_history = data["all_obs"][:, :, :2]      # [N, 8, 2] FPV noisy xy
clean_past    = data["all_pred"][:, :8, :2]     # [N, 8, 2] BEV clean past xy
clean_future  = data["all_pred"][:, 8:, :2]     # [N, 12, 2] BEV clean future xy
valid_mask    = data["all_obs"][:, :, 5]         # [N, 8] visibility mask

Dataset Details

EgoTraj-TBD

• Source: TBD dataset (CMU campus, 17 recording sessions)
• Perception: YOLOv8 detection + BotSort tracking on real ego-view video
• Matching: Hungarian algorithm with weighted MSE (location + velocity + acceleration)
• Sampling: 2.5 fps (stride 12 at 30fps source), 8-frame obs (3.2s) + 12-frame pred (4.8s)
• Statistics: 36,947 sequences, FPV noisy rate 0.37, history MSE 0.66m

T2FPV-ETH

• Source: T2FPV simulated ego-centric noise on ETH-UCY
• Folds: eth, hotel, univ, zara1, zara2 (leave-one-out evaluation)
• 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.