search

princetonvisualai / geode_dataset

3 stars 2 forks source link

Image country labels seem wrong #1

Open polvanrijn opened 1 year ago

polvanrijn commented 1 year ago

Dear creators of the dataset,

This is a great initiative! However, I worry the country labels in the published dataset are off.

I can say -- living in Germany for more than 10 years now -- that the majority of images are definitely not taken in Germany… Germany_bus_55921 Germany_waste_container_42793 Germany_road_sign_44355

You can clearly identify them not being in Germany by the car and street signs.

Similar problems exist in France, for example, "West Midlands" is in the UK and not in France. Also, you can clearly see the currency pounds on the bus… France_bus_61003

So how did you check if the images were taken in a given country? How did you make sure the images were really made by the participant and not copied from the internet (is this guaranteed by using the Appen Mobile app)? Did you store the IP address of the participant? Or use GPS data in the images?

Thanks for replying to my questions.

All the best, Pol van Rijn

vramaswamy94 commented 1 year ago

Hi, Thanks for bringing this to our attention. We do have GPS data for the images, and are working to ensure that the locations are correct, but it does appear that there are some errors within the labeling (our estimate, based on the images we checked so far is < 1%). We will let you know once we fix it!