Hugging Face's logo

Datasets:
doppelgangers25
/
VisymScenes

ArXiv:
License:
Dataset card Files Community
Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

VisymScenes Dataset

The VisymScenes Dataset is a worldwide video dataset collection for 3D modeling. We tasked freelancers around the world to select up to 8 interesting sites near them containing a man-made structure, then collect 32 or more videos walking around each site. This dataset is curated using the Visym Collector platform for privacy preserving visual AI data collection.

VisymScenes contains:

  • 258104 images from 149 sites collected from 42 cities and 15 countries
  • Data collected during the period July 2023 - March 2024
  • Video frames collected from ground-level mobile devices in first-person perspective, outdoors, in daylight, while walking around sites
  • 140 sites revisited over hours, days and months
  • Blurred personally identifiable information of all people and vehicles

Directory Structure

The directory structure is organized into subdirectories of sites, collectors and timestamps.

README.md                                              
{SITE_ID}/                                             
  {COLLECTOR_ID}-{DEVICE_TYPE}/                        
    {TIMESTAMP}/
      {SITE_ID}-{COLLECTOR_ID}-{DEVICE_TYPE}-{TIMESTAMP}-NNNNNN.jpg   
      {SITE_ID}-{COLLECTOR_ID}-{DEVICE_TYPE}-{TIMESTAMP}-NNNNNN.json

Dataset statistics

Images = 258104
Videos = 6368
Countries = 15
Cities = 42
Sites = 149
Sites revisited at least three times = 140
Intrinsics rate = 0.569951
IMU rate = 0.858700 Median GPS accuracy (1-sigma, meters)={'lat': '7.81', 'lon': '7.81'}
Device OS = 2
Top-14 Device types = ['iPhone 11 Pro', 'SM-A325F', 'CPH2387', 'V2120', 'iPhone 11 Pro Max', 'MAR-LX2', 'TECNO KG7h', 'CPH1933', 'V2254', 'iPhone XS', 'Redmi 8A', 'M2101K6G', 'Redmi Note 9 Pro', 'iPhone 7']
Collectors = 26
Collectors with >10 submissions = 24
Collectors with >100 submissions = 19
Cities (by collection)=[('Nairobi, Nairobi County, Kenya', 1187), ('Tagum, Davao del Norte, Philippines', 1073), ('Buenos Aires, Autonomous City of Buenos Aires, Argentina', 588), ('Hafizabad, Punjab, Pakistan', 556), ('Shah Alam, Selangor, Malaysia', 423), ('Dera Ghazi Khan, Punjab, Pakistan', 394), ('Caracas, Capital District, Venezuela', 322), ('Kajang Municipal Council, Malaysia', 310), ('Paschim Vihar, Delhi, India', 296), ('Dubai, Dubai, United Arab Emirates', 285), ('Gulistan-e-Jauhar, Sindh, Pakistan', 232), ('Karachi Division, Sindh, Pakistan', 167), ('Woodley/Kenyatta/Golf Course ward, Kenya', 140), ('Deira, Dubai, United Arab Emirates', 126), ('Ndenderu ward, Kiambu, Kenya', 105), ('Rostock, Mecklenburg-Vorpommern, Germany', 104), ('Lexington Massachusetts, United States', 102), ('Gulshan Town, Sindh, Pakistan', 99), ('Montevideo, Montevideo, Uruguay', 78), ('Woburn, Massachusetts, United States', 71), ('Copenhagen, Capital Region of Denmark, Denmark', 67), ('Kajang Municipal Council, Selangor, Malaysia', 55), ('Business Bay, Dubai, United Arab Emirates', 42), ('Stockholm, Sweden', 40), ('Punjab, Punjab, Pakistan', 37), ('Muar, Johor, Malaysia', 35), ('Ongata Rongai ward, Kajiado County, Kenya', 35), ('Ruaka, Kiambu, Kenya', 35), ('Kaputiei North ward, Kenya', 32), ('Malakisi/South Kulisiru ward, Bungoma County, Kenya', 31), ('Qalander Pura, Punjab, Pakistan', 30), ('Gdynia, Pomeranian Voivodeship, Poland', 20), ('Shadipura, Punjab, Pakistan', 13), ('Lahore, Punjab, Pakistan', 10), ('Johar Town, Punjab, Pakistan', 5), ('Islamabad, Islamabad Capital Territory, Pakistan', 3), ('Stavanger, Norway', 2), ('Bahía Blanca, Buenos Aires, Argentina', 1), ('Gothenburg, Sweden', 1), ('Matathani ward, East Kenya, Kenya', 1), ('Nairobi Central ward, Kenya', 1), ('Putrajaya, Putrajaya, Malaysia', 1)]
Detected transients: {'license plate': 32448, 'vehicle': 112811, 'face': 140232, 'person': 142586}
Transient rate (at least one detected object per frame): 0.391

Metadata

All metadata files are exported in an open JSON format that includes the following information:

  • GPS accuracy per frame reported as one-sigma lat/lon/alt uncertainty in meters
  • Orientation quaternion per frame
  • Intrinsic calibration reported by manufacturer
  • Bounding boxes for transient occlusions (people, vehicles)
  • Bounding boxes for removed PII (faces, license plates)
  • Device type, operating system (OS) and OS version
  • Universal Unique ID (UUID) for each site, collector, video, transient occlusion and removed PII
  • Video upload codec/bitrate/framerate/timestamp
  • Device orientation state (left landscape, right landscape, portrait)
  • Nominatim GPS geolocation
  • VIPY JSON field for dataset visualization and transformation
  • Video upload timestamp (UTC)
  • Mobile app version
  • 85.9% of the submissions were collected by devices that have gyroscopes. Submissions without gyroscopes have "None" reported for the exterior orientation angles (omega, phi, kappa).
  • 57.0% of the submissions were collected by devices where the manufacturer report calibrated camera intrinsics. Submissions without calibration have intrinsics that default to zero lens distortion coefficients, image centered principal point, and focal length equal to the video width. The field "intrinsic_calibration" is set to None if there are no reported intrinsics from the device manufacturer, which can be used to filter these default values.

Export parameters

  • Resolution: 1080x1920
  • Export framerate: 1Hz
  • Export JPEG quality: 0.9
  • Device orientation: portrait, right/left landscape (collector dependent)
  • Source video bitrate: 8285438
  • Source video duration: 40 seconds
  • Source video framerate: 10-30Hz (device dependent)
  • Source video codec: H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10
  • Source video pixel_format: yuvj420p

Collection Procedure

All collections are made using the Visym Collector platform as described in:

J. Byrne, G. Castanon, Z. Li and G. Ettinger, “Fine-grained Activities of People Worldwide”, Winter Conference on Applications of Computer Vision (WACV), 2023 [arXiv:2207.05182]

The campaign workflow is as follows:

  • Collectors are recruited and onboarded with a summary of the data we will collect, how the data will be used, stored and shared, and our payment terms
  • Collectors submit an example video of a candidate site near them
  • Candidate sites are checked by our review team and are approved or rejected
  • Approved sites are authorized for submission of additional ground level collections
  • Approved site collections may include additional requirements (e.g. illumination, weather, viewpoint)
  • Submitted videos are monitored live on the admin dashboard and checked by our review team
  • Payments are made bi-weekly to our collector community for accepted videos
  • Accepted videos are exported to frames and processed for PII redaction of faces and license plates
  • Redacted frames are checked by our review team to confirm PII removal
  • Redacted frames are delivered in JSON format

Recruitment Invitation

Here are the instructions that are given to collectors by email when they are invited to contribute to this campaign:

Hello,

You have been invited to contribute to a new campaign in the Visym Collector app. This campaign will be collecting videos of sites for 3D modeling. Our goal is to collect many short videos of different large buildings around the world so that we can create a 3D "fly through" of a site. You are responsible for choosing up to 8 sites near you that have a large man-made structure. Use the "Site Selection" project in the app and send us an example video as you walk around the outside of the building of each site. After review, we will request up to 32 additional videos of each site. The videos should be recorded with you walking around the site, like you are collecting a video of a local landmark to show a friend.

Here are two examples:

Walking around the outside of a large library

Walking around a parking lot

Campaign requirements:

  • Sites should be large, approximately 200m x 200m outdoors.
  • Sites should be interesting, containing at least one man-made structure, like a building or a local landmark.
  • Sites must be separated by at least 200 meters.
  • You must walk around the site while collecting your videos, and each one must be at least 40 seconds long.
  • Each video should be collected walking around a new place in the site. Our goal is to view this site from the back, front and sides, close up and far away. Try to get the whole building in view.
  • You may make up to 8 submissions of new sites. We will review your submissions and approved sites will be authorized for up to 32 submissions walking around different places in this site.
  • Once a site is approved, you will receive a new collection to submit more videos at this site with a helpful map link to remind you where it is.
  • There is no rate limit for this campaign. You may submit all videos in a single day, but site reviews may take up to 24 hours.
  • We will request many more sites and submissions from collectors over time. This campaign will run for 18 months.
  • It is OK if there are some people or cars in the background, but please try to minimize them, if possible. We will blur them before sharing your videos with others.
  • You should only collect videos in safe, publicly accessible locations. Please do not collect at sensitive government sites.
  • You are required to upgrade to the latest version (2.8.1) in the Android and iOS app stores.

All submissions will be made using the Visym Collector mobile app. The current active collections that can be submitted will be shown in-app under "Projects". The price we pay for each accepted video is shown in the collection description. Your submissions will be available to review within the app under "Videos". The submissions will also be available in your Collector Dashboard which shows the collection status from the review team. Your dashboard will contain a summary of the payments planned for the week, which you can use to track your expected totals, and see the response from our review team if a video was not accepted. Payments are made through Visym Labs, and are paid every two weeks by 5pm ET.

Here are instructions to get you started with the Visym Collector app:
https://visym.com/quickstart

Here is a link to your dashboard. You may log in using the same account used for the app. We provide payments via Payoneer or PayPal. You should update your payment details under "Profile" in the dashboard link so we can pay you. We are targeting a $30 USD / hour rate. Initially, you will submit 256 videos, then 256 videos every few months for a total of approximately 1600 videos.
https://collector.visym.com

Please contact us at [email protected] with any questions, problems or payment concerns.

We look forward to working with you.

Thanks, -Collector Team

Site Selection Instructions

Here are the instructions that are presented to the freelancer workforce in the Collector mobile app for selecting sites:

Collect an example video of a new site of at least 40 seconds. You should walk around the site, holding your device to record while walking. You may turn your device left, right, up or down to capture interesting features of the site while you walk, like you are collecting a video to show a friend.

Here is one example walking around the outside of a large library. Try to avoid foliage if you can. Your video should look like this.

Here is a second example walking around a parking lot. You may hold your device either in landscape or portrait orientation.

Please note that we do not want videos containing faces or license plates. It is OK if there are some people or cars in the background, but please don't record them if possible. We will blur them out after we receive your video.

Each site you choose must be separated by at least 200 meters. After review, we will send you a request for more videos in each approved site. Each approved site will be available to you as new collection along with a helpful map link and approximate address to remind you where it is.

You may submit at most 8 example sites for us to review. There is no rate limit for this campaign, so you may submit all videos in a single day if you choose.

Video Collection Instructions

Here are the instructions that are presented to the freelancer workforce in the collector mobile app for collecting videos in approved sites. This collection was auto-generated for a collector in Sweden who submitted a site in Stockholm.

Collect at most 32 videos, each at least 40 seconds walking around a 200 meter x 200 meter site centered at this map location:

10, Galärvarvsvägen, Djurgårdsstaden, Djurgården, Östermalms stadsdelsområde, Stockholm, Danderyds kommun, Stockholm County, 115 21, Sweden

You should walk around this site, holding your device to record while walking. You may turn your device left, right, up or down to capture interesting features of the site while you walk, like you are collecting a video to show a friend. We want all views of this site, from the sides, front, back, near and far. Each collection must be a different view of this site, and may include multiple man-made structures in this 200m x 200m area. Note that the collection name we provide may not be exactly where you collected, but it should be near by. Use the map link above to remind yourself precisely where you collected.

Here is one example walking around the outside of a large library. Try to avoid foliage if you can. Each video must be from a different viewpoint and location within this site.

Here is a second example walking around a parking lot. You may hold you device either in landscape or portrait orientation. Make sure each video is from a different location within this 200m x 200m site.

There is no rate limit for this collection. You may submit all 32 video in one day if you choose, but each video submission must be from a new place within this site, so that all 32 videos together capture the entire site.

Please note that we do not want videos containing faces or license plates. It is OK if there are some people or cars in the background, but please don't record them if possible. We will blur them out after we receive your video."

Submission Verification

Each submission from the worforce is reviewed to confirm:

  • The video contains at least one man-made structure
  • The video has the required metadata
  • The video was collected in the required conditions (daylight, outdoors)
  • The video was collected at the required 200m x 200m site
  • Sites are not repeated across collectors in the same city
  • Sites are publicly accessible and not sensitive to national/state/local/tribal governments

Accepted videos are paid out to the collectors biweekly at an agreed upon rate per accepted video.

Personally Identifiable Information (PII) Redaction

All license plates and faces are blurred in exported images prior to release. Object detection for redaction is performed at a lower confidence threshold (0.3) than dataset 1, and includes motorbikes as a vehicle class.

Transient Occlusions

To access the bounding boxes for transient occlusions in python:

pip install vipy

import vipy, json

d = vipy.util.readjson('/path/to/NNNN.json')
im = vipy.image.Scene.from_json(json.dumps(d['vipy']))
im.objects()

Release History

  • June 7, 2024 (Initial release)
  • Dec 6, 2024 (arXiv release)

Acknowledgement

Supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0035. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government.

License

Creative Commons Attribution 4.0 International (CC BY 4.0).
https://creativecommons.org/licenses/by/4.0/

Contact

Visym Labs <[email protected]>

Downloads last month
56