tkuri / papers

Summarizing the papers I have read (Japanese)
41 stars 1 forks source link

ImagePairs: Realistic Super Resolution Dataset via Beam Splitter Camera Rig #80

Open tkuri opened 4 years ago

tkuri commented 4 years ago

論文概要

今までのDNNでの超解像やNRなどの失敗要因は手法の良し悪しじゃなくて結局「データが少ないから」ではないかという問題提起から、低解像度と高解像度のペアの大規模データセット(11421枚)をビームスプリッターを利用した治具で構築した。Microsoft。6/15に公開予定。

既存のSOTAの手法らの大幅な性能向上を確認。LRとSRのカメラ間でのレンズ違いによる様々なずれは局所的にアライメントをして解決する。CVPR2020 Workshop。

bib_20200429 00

感想

現像後の画像だけでなくRAWも同時に公開する予定なのでデモザイクやホワバラ、トーンマッピングなどカメラ信号処理の開発にも非常に有用だろう。楽しみ。

UESTCrookieLI commented 2 years ago

Hi,@tkuri It is very excited that i meet someone like you who haved read this paper too. I have a doubt that why the color of HR and LR in this paper are different. image I can not use this dataset to train a good deep learning network due to the color is different in HR and LR.

someone said it is because of Polarization of light, but I do not think so.(I know there is Neutral beam splitter that will not change the color)

Do you have some thought solve my problem?

If you reply me, i will very gald.

Look forward to you reply!