Metadata for the 10k dataset

[hutec]

Hi,

thanks for a very nice dataset and this repo!

I'm wondering if the metadata attributes (weather, scene and timeofday) are available for the 10k datasets?

[fyu]

It is not available to all the images for the segmentation dataset yet.

[sh-jj]

Hi, such a dataset with rich attribute information is really helpful!
Ummm, my question is how to download the meta-information (for IMAGE 100K data). I have searched a long time, but I still do not find which file I need to download on the download page (https://bdd-data.berkeley.edu/portal.html#download). 😂
Could you please tell me which file contains the meta-information (weather, time, scene)?
Thanks!

[fyu]

The attributes can be downloaded by the Label button https://doc.bdd100k.com/download.html#labels

[sh-jj]

The attributes can be downloaded by the Label button https://doc.bdd100k.com/download.html#labels

Thanks again！ It's really helpful for my research.

[SnowdenLee]

Hi @fyu,
is there any update about releasing the meta information for 10K dataset? Thanks a lot in advance!

[Nimra28]

Can any one tell me the PC requirements to work on this dataset? @fyu

[HaoLin97]

@fyu Is there any update on this?

[Nimra28]

Can any one tell me the PC requirements to work on this dataset? @fyu

[okaynils]

Using available Metadata for the 10k Dataset

As @fyu has mentioned, metadata is not directly available for the 10k dataset.

I got curious since I am currently taking a deep learning class in which I train and compare semantic segmentation models to take a closer look at the impact types of scenes in self-driving environments have.

It looks to me like the 10k dataset contains sampled images from the 100k dataset so when I tried to match the image filename references in the det_train.json and det_val.json to the 10k images filenames I found that there is some overlap. Of the 10k images 3426 have an annotated reference in either the det_train.json or det_val.json. All the matches reference to images inside the train-partition of the 10k dataset. So not all 10k images are annotated thus only the 3426 of 8000 train-images can be accessed with existing attributes ("scene", "weather" and "timeofday").

Note: Again, since this overlap only has references to the train-partition, anyone interested in using the metadata would need to partition the dataset themselves.

You can find the det_train.json and det_val.json available to download at https://dl.cv.ethz.ch/bdd100k/data/.

[wang1528186571]

Has the conflict been resolved in the json in the link above? If not, could you provide a resolved version?

[okaynils]

Has the conflict been resolved in the json in the link above? If not, could you provide a resolved version?

@wang1528186571 I do not have a resolved .json but I have made a PyTorch Dataset class that only loads in samples that have scene metadata: https://github.com/okaynils/fhnw-ds-dlbs/blob/main/data/bdd100k_dataset.py