Princeton Visual AI Lab                     Meta AI

GeoDE: Crowd-sourcing details

 

Crowdsourcing GeoDE

We ask participants from 6 different regions of the world to send us images of 40 different objects. This results ins a dataset comprising of 61,940 images roughly balanced across both these objects and regions.Instructions given to the workers is shown in the following image.


General Instructions
In this task, you will submit up to 3 photos of the same type of object (e.g., upload 3 photos of 3 different bags;
please do not upload 3 photos of the same bag from different angles).
1. Please make sure the location function is enabled for the camera.
2. The photo resolution should be at least 640 x 480.
3. All images should be new photos captured with Appen Mobile.
4. Please make sure it’s a single object per image.
5. Please make sure it’s a well-lit environment and the object is clearly visible in the photos.
6. Please make the object occupy at least 25% of the image.
7. Objects captured are foregrounded and not occluded.
8. Objects should not be blurred, e.g., motion blur.
9. No effects or filters added (cropping is acceptable).
10. Please try to avoid capturing people in the images (it’s OK if people are blurry in the background and far
from the camera).
11. Please try to avoid capturing vehicle license plates in images


The final dataset comprises of images from over 4500 workers within these different countries, representing a range of genders, ages and races, as shown in the image below.

centered image


Objects and regions in GeoDE

In the following two tables, we list the objects collected, as well as the 6 world regions targeted when collecting GeoDE.
List of objects present in geode, loosely organized

Indoor common     bag, chair, dustbin, hairbrush/comb, hand soap, hat,
light fixture, light switch, toothbrush, toothpaste/
toothpowder

Indoor rare     candle, cleaning equipment, cooking pot, jug, lighter,
medicine, plate of food, spices, stove, toy

Outdoor common     backyard, car, fence, front door, house, road sign,
streetlight/lantern, tree, truck, waste container

Outdoor rare     bicycle, boat, bus, dog, flag, monument, religious building,
stall, storefront, wheelbarrow



GeoDE consists of images from six world regions. Within each region, there are 3-4 countries contributing to most of the images, chosen to balance the diversity of the images against practical data collection considerations.

West Asia              Saudi Arabia, United Arab Emirates, Turkey

Africa     Egypt, Nigeria, South Africa

East Asia     China, Japan, South Korea

Southeast Asia     Indonesia, Philippines, Thailand

Americas              Argentina, Colombia, Mexico

Europe              Italy, Romania, Spain, United Kingdom



Contact

Vikram V. Ramaswamy (vr23@cs.princeton.edu)


back to main page