Question about feature transferability compared to supervised approaches

wvangansbeke / Unsupervised-Semantic-Segmentation

Unsupervised Semantic Segmentation by Contrasting Object Mask Proposals. [ICCV 2021]

https://openaccess.thecvf.com/content/ICCV2021/html/Van_Gansbeke_Unsupervised_Semantic_Segmentation_by_Contrasting_Object_Mask_Proposals_ICCV_2021_paper.html

Other

398 stars 54 forks source link

Question about feature transferability compared to supervised approaches #26

Closed GerardMaggiolino closed 2 years ago

GerardMaggiolino commented 2 years ago

Were any experiments ran to evaluate how well representations learned by the model transfer to different datasets under linear eval in comparison to features learned by fully supervised counterparts? This is a nice property of self supervised image level representations, so I’m wondering if the same applies here.

Awesome paper btw, looking forward to using it in my work 🙂

wvangansbeke commented 2 years ago

Hi @GerardMaggiolino,

Thank you for the kind words. Yes. Transfer results from ImageNet to PASCAL, COCO and DAVIS can be found in Table 5 of the paper.