Dear ADE20K team,
First, thank you for maintaining and openly sharing such a valuable dataset for the vision community. I am currently using ADE20K for research purposes, and while exploring the annotations, I noticed a few typographical inconsistencies in some object names that may affect label consistency and downstream analyses.
For example, the following entries appear to correspond to the same concept (“school desk”) but with different spellings:
| name |
name_ndx |
| sckool desk |
3196 |
| scool desk |
3197 |
| sscool desk |
3198 |
| schoo desk |
3199 |
These seem to be minor typographical variants that could be unified under the correct label "school desk" to improve consistency in object-level statistics and model training.
Could you please confirm whether there is a preferred channel or contribution process to help identify and correct these inconsistencies (e.g., through a pull request, structured report, or CSV correction proposal)?
I would be glad to assist in compiling a list of similar issues across the dataset if that would be helpful.
Thank you for your time and for your outstanding work on ADE20K.