knowledge distillation

[xwu6614555]

Hello, thanks for your work!

The paper mentions using the Pro model to distill the Edge model to improve performance. Specifically, how much improvement does it achieve compared to methods without distillation?

[KitWong2048]

Hello, thanks for your work!

The paper mentions using the Pro model to distill the Edge model to improve performance. Specifically, how much improvement does it achieve compared to methods without distillation?

Thank you for your interest. We conducted an ablation study on a small dataset consisting of O365, Flickr, and Visual Genome (VG). The experimental results demonstrate that models trained with distillation outperform those trained without distillation, achieving an improvement of approximately 2 percentage point on the COCO dataset.