search

Linwei-Chen / LIS

IJCV2023 Instance Segmentation in the Dark
MIT License
73 stars 6 forks source link

problems about experiment settings in the paper #19

Open beyondbatman-master opened 2 weeks ago

beyondbatman-master commented 2 weeks ago

Your work is very interesting, and I have two questions regarding the experiments in your paper:

  1. In the paper, you compared the approach of using image enhancement followed by object detection (such as HE, GLADNet, Retinex-Net, EnlightenGAN, Zero-DCE, SID, and REDI). Were these methods using the checkpoints provided by their original papers, or were they retrained on the noisy COCO dataset or trained from scratch?

  2. The Enhance + Denoise and Integrated Enhance + Denoise methods use enhancement networks for RGB images and RAW images(HE, GLADNet, Retinex-Net, EnlightenGAN, Zero-DCE for RGB&SID, and REDI for RAW), respectively. During inference, are the enhancement networks still conducted on images in the RGB domain and RAW domain, respectively?

Looking forward to your reply.

Linwei-Chen commented 2 weeks ago

Thanks for your interest!

  1. We prefer to use its original checkpoint.
  2. Yes, we use their oirignal input domain.