discusses shortcomings in the famous ImageNet dataset and the authors have recomendations for better processes.
There is a difference in that imagenet seems to aim at a single label per image, they mention a shortcoming that "many imagenet images have seveal objects" - but for the type of annotation we're going for here (eg supporting segmentaton) that's standard.
Offsensive labels for people in the wordnet database were a specific problem (which does seem to validate the curated label list idea), and they want to move to upgrade the variety they store.
I thought it was an interesting article generallyt that might point at fresh space and future direction for image datasets.
https://venturebeat.com/2020/07/15/mit-researchers-find-systematic-shortcomings-in-imagenet-data-set/?fbclid=IwAR06Z4DSrNd54dWcHCVOLt3DAn8cjbHYukB8UT5Hu8XNCi58nkiUvUeMSAA
discusses shortcomings in the famous ImageNet dataset and the authors have recomendations for better processes. There is a difference in that imagenet seems to aim at a single label per image, they mention a shortcoming that "many imagenet images have seveal objects" - but for the type of annotation we're going for here (eg supporting segmentaton) that's standard. Offsensive labels for people in the wordnet database were a specific problem (which does seem to validate the curated label list idea), and they want to move to upgrade the variety they store.
I thought it was an interesting article generallyt that might point at fresh space and future direction for image datasets.