uhhyunjoo
commented
2 years ago Evaluation metrics
data, image, feature, perception quality metrics
- peak signal-to-noise ratio (PSNR) : to evaluate quality of synthesized volumes at the data level
- structural similarity index (SSIM) : to evalutae the quality of volume rendering images at the image level
besides volume rendering, we also compare against BI in terms of volumetric features, expressed in the form of isosurface
-
to quantiffy the similarity between two isosurfaces extracted, from the synthesized and GT volumes, we compute their isosurface similarity (IS) at the feature level.
-
mean option scroe (MOS) : to evaluate how close its rendering image is with respect ot the GT image. -> we recruited ten students and asked them give a five-point score, 1 (loweset perceived quality) ~ 5 (hightest perceived quality) : the mean of these scores is reported as MOS
SSR-TVD 를 BI와 CNN 에 대해서 비교
- data level (PSNR)
- image level (SSIM)
데이터셋
- five jets
- vortex
- ionization (He+)
- ionization (H+)
- one variable X of one data se used for training,
- another variable Y of the same data set is applited for inference
- X -> Y
adversarial learning 을 사용해서, time-varying data (TVD) 에 대한 spatial supter-resolution (SSR) 을 생성하는 새로운 딥러닝 프레임워크
low-resolution sequence 인 128x128x128 이 input 으로 주어지면, ouput 으로 high-resolution sequence 인 512x512x512 를 생성하는 네트워크
low-resolution volume sequence 에서 high-resolution volume sequence 를 생성하는 것은 두 가지 과제가 있는데,
high-resolution sequence should maintain good temporal coherence
human perception should be considered in the evaluation of upscaled volumes.
이런 challenges 를 해결하기 위해, 해당 논문에서는 딥러닝 접근 방식을 사용해서, neural network 가 low-resolution volumes 로부터 features 와 그들의 relationships 를 배울 수 있고, 따라서 high-resolution volumes 를 high-wuality 로 생성할 수 있다고 한다.
sparse coding framework 와 image super-resolution technique 에 영감을 받아, 본 논문에서는 SSR-TVD를 제안한다. a generative model 과 two discriminators 를 갖는 그들의 세 가지 다른 losses를 사용해서, sequence of volumes 에 대한 spatiotemporally coherent of SSR 을 생성하기 위해
adversarial loss, content loss, feature loss 를 다 고려하는 loss functino 을 최소화해나감으로써 SSR-TVD 를 학습시킴
our approach 의 effectiveness 를 보여주기 위해, several time-varying data sets of different characteristics 를 이용하여 quantitative and qualitative results 를 보여준다.
본 논문의 SSR-TVD 와 widely-used BI 랑 a solution soley based on CNN 를 비교한다.
data, dimage, feature, perception levels 에 대해 더 나은 volume quality 를 달성했다.
third row of Fig.4 shows our SSR-TVD results
original volumes 를 downsampling 해서 low-resolution volumes 를 얻고, inference 할 때 SSR-TVD가 original resolution 으로 volumes back 한다.
The contributions of this paper