WTS Dataset

WTS Dataset Overview

Dataset Introduction

• Name: WTS: A Pedestrian‑Centric Traffic Video Dataset for Fine‑grained Spatial‑Temporal Understanding
• Source: Woven by Toyota, Inc.
• Purpose: Emphasize detailed vehicle and pedestrian behaviors in various traffic events, including accidents.
• Content: Over 1.2 k video events spanning more than 130 distinct traffic scenes, combining vehicle‑ego and fixed‑high‑angle camera views.
• Additional Resources: Approximately 4.8 k pedestrian‑focused traffic videos from BDD100K with detailed textual annotations.

Dataset Structure

• Video Data: Stored in the videos folder, containing both real‑world WTS videos and the pedestrian‑centric subset (BDD_PC_5K) filtered from BDD100K.
• Annotations: Provide BBox annotations for target pedestrians and vehicles, plus detailed scene descriptions covering position, attention, action, and context.
• Future Updates: Will add 3D gaze and position annotations.

Data Preparation

• BBox Annotations: Frame‑level annotations are supplied; a script frame_extraction.py is provided to extract frames matching annotation IDs.
• Shared Annotations: Videos within the same scenario folder share the same top‑level annotations.

Evaluation

• Validation Set: A validation split for video‑to‑text tasks is provided, mirroring the training folder structure under val.
• Submission Format: For AI City Challenge 2024 Track 2, submissions should be a JSON file containing predictions for all test videos.

License and Citation

• License: Refer to the WTS homepage for licensing details.
• Citation: If the dataset benefits your research, please cite the associated publication.

Update List

• Completed: View list, first‑frame BBox annotations, generated BBoxes, evaluation code.
• Pending: 3D gaze annotations, dataset arXiv paper.