WTS:A Pedestrian-Centric Traffic Video Dataset for Fine-grained Spatial-Temporal Understanding
Dataset download link (serves as official source for AI City Challenge 2024 Track2 @ CVPR2024)
-
[2024/07/25]WTS dataset is accepted by ECCV24! Updated arXiv paper! -
[2024/07/17]3D Gaze annotation is released! It is available in the previously provided annotations in dropbox or google drive link (keep updating more). -
[2024/06/30]AI City Challenge 2024 is finished this year! Please directly require the WTS dataset from this form -
[2024/02/23]updated the evaluation script for AIC Track2 evaluatin system submission. -
[2024/02/22]AIC challenge Track2 Test data released and evaluation is open ! -
[2024/02/22]BBox of whole train/val of dataset are updated! Please download it from provided dropbox or google drive link to your email. -
[2024/01/31]updated with the evaluation script for AIC Track2. -
[2024/01/22]AI City Challenge 2024 Track2 opened. -
[2024/01/22]WTS dataset released. -
[2024/01/21]WTS dataset project page released.
- List of views used as main reference during annotation. (2024/01/23)
- First frame annotated bbox for each segment. (2024/02/22)
- Generated bbox for target pedesrian and vehicle. (2024/02/22)
- Evaluation code. (2024/02/23)
- 3D Gaze annotation release. (2024/07/17)
- Dataset arXiv paper.
The Woven Traffic Safety (WTS) Dataset from Woven by Toyota, Inc., is designed to emphasize detailed behaviors of both vehicles and pedestrians within a variety of staged traffic events including accidents. Comprising over 1.2k video events across over 130 distinct traffic scenarios, WTS integrates diverse perspectives from vehicle ego and fixed overhead cameras in a vehicle-infrastructure cooperative environment. Each event in WTS is enriched with comprehensive textual descriptions of the observed behaviors and contexts. For diverse experimental purposes, we also provide the same detailed textual description annotations for approximately 4.8k publicly sourced pedestrian-related traffic videos from BDD100K for external use as training/test resources, etc.
WTS provides the largest number of videos with long fine-grained video descriptions with 3D spatial information in the traffic domain.
we have two kinds of video data provided for use:
- Collected real-world WTS data including traffic accidents.
- Filtered pedestrian-centric videos from BDD100K data with our WTS annotations as
BDD_PC_5K
Video Data
All collected WTS video data are stored under the videos folder.
For the multiple-view videos, part of the scenarios will proceed across the views.
We also provide the bbox annotations for the target pedestrian and vehicle for selecting the target video once needed.
videos
├── train
│ ├── 20230707_12_SN17_T1 ##scenario index
│ │ ├── overhead_view ##different overhead view about the scenario
│ │ │ ├── 20230707_12_SN17_T1_Camera1_0.mp4
│ │ │ ├── 20230707_12_SN17_T1_Camera2_3.mp4
│ │ │ ├── 20230707_12_SN17_T1_Camera3_1.mp4
│ │ │ └── 20230707_12_SN17_T1_Camera4_2.mp4
│ │ └── vehicle_view ##vehicle ego-view about the scenario
│ │ └── 20230707_12_SN17_T1_vehicle_view.mp4
│ ├── 20230707_15_SY4_T1
│ │ ├── overhead_view
│ │ │ ├── 20230707_15_SY4_T1_Camera1_0.mp4
│ │ │ ├── 20230707_15_SY4_T1_Camera2_1.mp4
│ │ │ └── 20230707_15_SY4_T1_Camera3_2.mp4
│ │ └── vehicle_view
│ │ └── 20230707_15_SY4_T1_vehicle_view.mp4
...
All pedestrian-related videos from BDD are stored under external folder:
external/
└── BDD_PC_5K
├── videos
│ ├── train
│ ├── video1004.mp4
│ ├── video1006.mp4
│ ├── video1009.mp4
│ ├── video100.mp4
│ ├── video1015.mp4
...
Annotation
Two kinds of annotations now are available.
- BBox for the target pedestrian and vehicle.
- Description for the traffic scenario focuses on the
location, attention, behavior, contextregarding the pedestrian and vehicle. - 3D Gaze of the pedestrian
Will update the 3D Gaze and Location annotations for use (stay tuned).
Notice that the videos from overhead and vehicle in the same scneario index folder will share the same caption.
annotations
├──caption/
├── train
│ ├── 20230707_12_SN17_T1 ##scenario index
│ │ ├── overhead_view
│ │ │ └── 20230707_12_SN17_T1_caption.json ##caption shared by multiple views
│ │ └── vehicle_view
│ │ └── 20230707_12_SN17_T1_caption.json ##caption is the same as overhead view, only timestamp is for vehicle view for finding the same segment.
│ ├── 20230707_15_SY4_T1
│ │ ├── overhead_view
│ │ │ └── 20230707_15_SY4_T1_caption.json
│ │ └── vehicle_view
│ │ └── 20230707_15_SY4_T1_caption.json
...
Caption JSON format:
{
"id": 722,
"overhead_videos": [ ## caption related videos
"20230707_8_SN46_T1_Camera1_0.mp4",
"20230707_8_SN46_T1_Camera2_1.mp4",
"20230707_8_SN46_T1_Camera2_2.mp4",
"20230707_8_SN46_T1_Camera3_3.mp4"
],
"event_phase": [
{
"labels": [
"4" ##segment number
],
"caption_pedestrian": "The pedestrian stands still on the left side behind the vehicle, ...", ##caption for pedestrian during the segment
"caption_vehicle": "The vehicle was positioned diagonally to ...", ##caption for vehicle during the segment
"start_time": "39.395", ##start time of the segment in seconds, 0.0 is the starting time of the given video.
"end_time": "44.663" ##end time of the segment in seconds
},
...
BBox annotations for the first frame of each segment in the video, notice that not all videos have the bbox as there are some views that can not cover the pedestrian or vehicle.
annotations
├── bbox_annotated ##bbox for the first frame of each segment
│ ├── pedestrian ##bbox for pedestrian and vehicle respectively
│ │ ├── train
│ │ │ ├── 20230707_12_SN17_T1
│ │ │ │ └── overhead_view
│ │ │ │ ├── 20230707_12_SN17_T1_Camera1_0_bbox.json ##bbox for per video
│ │ │ │ ├── 20230707_12_SN17_T1_Camera2_3_bbox.json
│ │ │ │ ├── 20230707_12_SN17_T1_Camera3_1_bbox.json
│ │ │ │ └── 20230707_12_SN17_T1_Camera4_2_bbox.json
...
BBox annotations for the whole frames generated by the video object tracking method for each segment in the video:
annotations
├── bbox_generated ##bbox for the whole frame of each segment
│ ├── pedestrian ##bbox for pedestrian and vehicle respectively
│ │ ├── train
│ │ │ ├── 20230707_12_SN17_T1
│ │ │ │ └── overhead_view
│ │ │ │ ├── 20230707_12_SN17_T1_Camera1_0_bbox.json ##bbox for per video
│ │ │ │ ├── 20230707_12_SN17_T1_Camera2_3_bbox.json
│ │ │ │ ├── 20230707_12_SN17_T1_Camera3_1_bbox.json
│ │ │ │ └── 20230707_12_SN17_T1_Camera4_2_bbox.json
...
BBox format follows COCO format, you could use our frame_extraction script to reproduce the frames with image_id.
{
"annotations": [
{
"image_id": 904, ## frame ID
"bbox": [
1004.4933333333333, ## x_min
163.28666666666666, ## y_min
12.946666666666667, ## width
11.713333333333333 ## height
],
"auto-generated": false, ##human annotated frame
"phase_number": "0"
},
{
"image_id": 905,
"bbox": [
1007.1933333333333,
162.20666666666668,
12.946666666666667,
11.713333333333333
],
"auto-generated": true, ##generated bbox annotation for the frame
"phase_number": "0"
},
...
For BDD, each caption annotations correspond with one video for use. The annotation format is the same as our collected data caption annotations.
external/
└── BDD_PC_5K
├── annotations
│ ├── train
│ │ ├── video1004_caption.json
│ │ ├── video1006_caption.json
│ │ ├── video1009_caption.json
│ │ ├── video100_caption.json
│ │ ├── video1015_caption.json
...
3D Gaze is also provided for each camera for a given dates and the structure is shown below:
annotations
├── 3D_gaze
│ ├── train
│ │ ├── 20230922_1_SN2_T1
│ │ │ ├── 20230922_1_SN2_T1_192.168.0.11-1_gaze.json ## Gaze per camera video
│ │ │ ├── 20230922_1_SN2_T1_192.168.0.12-2_gaze.json
│ │ │ └── 20230922_1_SN2_T1_192.168.0.28-3_gaze.json
...
Gaze annotation follows the similar structure as BBox, as shown below. The gaze (x, y, z) is in overhead camera coordinates in OpenGL axis convention (x to the right, y up, and z backward). image_id refers to the frame number in the overhead video.
{
"annotations": [
{
"image_id": 0, ## frame ID
"gaze": [
0.7267333451506679, ## x
0.27087537465994793, ## y
-0.6312568142259175 ## z
],
},
{
"image_id": 1,
"gaze": [
0.7267333451506679,
0.27087537465994793,
-0.6312568142259175
],
},
...
- Our BBox annotation is frame-based, you could use the below script to extract the frame to align the ID in our annotations.
python script/frame_extraction.py
- The caption annotations are shared with the videos in the same scenario folders in WTS collected video part.
We provide a list of which camera views are mainly used as a reference during the annotation, thuss which views are covered well of the whole scenarios in
view_used_as_main_reference.csvas below format. Users could feel free to use it.
<Scene,Viewpoint1_video_name,Viewpoint2_video_name,Viewpoint3_video_name,...> ##maximum to 4 views
20230707_4_CN35_T1,20230707_4_CN35_T1_Camera3_0.mp4,20230707_4_CN35_T1_Camera4_1.mp4
20230707_8_SN46_T1,20230707_8_SN46_T1_Camera2_1.mp4,20230707_8_SN46_T1_Camera2_2.mp4
20230707_9_SN1_T1,20230707_9_SN1_T1_Camera2_1.mp4,20230707_9_SN1_T1_Camera2_2.mp4
We provide the validation set for the video2text task with a given segment duration.
Video and its GT of the validation set are stored following the same structure as train under val folders.
For our collected data in WTS, the inference results are required to be provided per scenario.
Users could feel free to use the multi-view videos in the same scenario folders for validation purposes,
as well as multi-view videos in train for training purposes.
For BDD_PC_5K, each video has its caption GT in train and val, and validation will be done per video.
Regarding AI City Challenge 2024 Track2, the evaluation script is provided under evaluation/, notice that only caption annotations will be used for the challenge.
Submission(model output) format is defined as:
{
"20230707_12_SN17_T1": [ ##scneario index for multiple view situations OR video name for single view of "BDD_PC_5K".
{
"labels": [ ##segment number, this is known information will be given
"4"
],
"caption_pedestrian": "", ##caption regarding pedestrian
"caption_vehicle": "" ##caption regarding vehicle
},
{
"labels": [
"3"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"2"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"1"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"0"
],
"caption_pedestrian": "",
"caption_vehicle: ""
}
]
"20231013_105827_normal_192.168.0.14_1_event_2": [ ##scneario index for multiple view situations OR video name for single view "BDD_PC_5K".
{
"labels": [ ##segment number, this is known information will be given
"4"
],
"caption_pedestrian": "", ##caption regarding pedestrian
"caption_vehicle": "" ##caption regarding vehicle
},
{
"labels": [
"3"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"2"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"1"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"0"
],
"caption_pedestrian": "",
"caption_vehicle: ""
}
]
"video3334": [ ##scneario index for multiple view situations OR video name for single view "BDD_PC_5K".
{
"labels": [ ##segment number, this is known information will be given
"4"
],
"caption_pedestrian": "", ##caption regarding pedestrian
"caption_vehicle": "" ##caption regarding vehicle
},
{
"labels": [
"3"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"2"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"1"
],
"caption_pedestrian": "",
"caption_vehicle": ""
},
{
"labels": [
"0"
],
"caption_pedestrian": "",
"caption_vehicle": ""
}
]
}
Note:
- Please submit one JSON file that includes results for ALL videos in the test set. Otherwise, you will get warning saying counting zero score for missing videos.
- In your submissin, please make sure all video segments that the dataset annotation specified have both "caption_pedestrian" an "caption_vehicle" keys existing. Otherwise, you will get KeyError saying the key is not found.
will be updated with the arXiv paper soon.
Please refer to our license from WTS dataset homepage
Please consider citing our paper if you find WTS dataset is helpful for your works.
@misc{kong2024wtspedestriancentrictrafficvideo,
title={WTS: A Pedestrian-Centric Traffic Video Dataset for Fine-grained Spatial-Temporal Understanding},
author={Quan Kong and Yuki Kawana and Rajat Saini and Ashutosh Kumar and Jingjing Pan and Ta Gu and Yohei Ozao and Balazs Opra and David C. Anastasiu and Yoichi Sato and Norimasa Kobori},
year={2024},
eprint={2407.15350},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2407.15350},
}