zenseact / zod

Software Development Kit for the Zenseact Open Dataset (ZOD)
https://zod.zenseact.com
MIT License
92 stars 13 forks source link

Accuracy of data with 'fog' labels. #30

Closed siddhantsom closed 8 months ago

siddhantsom commented 8 months ago

My supervisor and I have been trying to work with the ZOD dataset for the purpose of weather detection. We were using the RGB data, and as labels we had decided to use the 'scraped-weather' parameter from the meta data json file, because it has values such as 'clear', 'rain', 'snow' etc, which can be used as class labels. One of the class labels is 'fog', however it is not clear as to how they are classified, because they easily come under some of the other labels such as 'cloudy', 'rain' and 'snow'. There is no clear indication of fog, at least in RGB. fog

029193_golf_2021-01-20T11_27_50 115510Z 030057_golf_2021-01-10T10_24_18 004076Z

003520_oscar_2021-04-15T14_37_55 113569Z 005790_oscar_2021-04-15T14_18_19 330970Z

wljungbergh commented 8 months ago

Hi.

As (somewhat) indicated by the name of the field, the data is scraped from online resources in post-processing. This unfortunately leads to metadata that is not correct. This is a limitation but I hope that you can still make use of the dataset.

However, if you are able to more accurately classify the weather of each frame (including fog etc.,), we would very much like to add it to the dataset. If you do so, please contact us either here or at opendataset@zenseact.com