search

jaeheungs / rdf_depth_from_focus

Official implementation of depth from focus using the ring difference filter (RDF)
15 stars 3 forks source link

How did you measure quality? #1

Open ApostolAnt opened 4 years ago

ApostolAnt commented 4 years ago

Hi I read your article and I have a few questions. I will be happy if you answer them.

1) How did you measure the quality? You say this is an RMSE, but can you describe more details? Maybe the formulas? 2) In the article [30] (Analysis of focus measurement operators for a form from focus), the paper processing time is about 10 ms, but in your work it is at least 500 ms. Why is this happening? Maybe I lost my understanding?

fuy34 commented 3 years ago

Hi @ApostolAnt @jaeheungs , May I ask where I can download the dataset used in this paper and if they contain the ground truth depth?

jaeheungs commented 3 years ago

@ApostolAnt , Sorry for the late reply.

  1. RMSE (root mean squared error) is measured by the sqrt( average ( (depth error per pixel)^2 ) )
  2. I believe the difference may be from [30] counting the amount of time needed per image and I am counting the time required for 30 images in the demo, so I am getting around 300ms on my machine. I don't know where you are getting the time of 500ms, but if it is from running my code then it might be that your machine is not as fast as mine.

@fuy34 It seems as though the HCI 4D Lightfield dataset is no longer available and has been replaced by the 2016 dataset. They do contain the GT depth.

fuy34 commented 3 years ago

Hi @jaeheungs ,

Thank you for your reply.

I realize the HCI 2016 is a light field dataset. I am not familiar with the light field camera. But in their ACCV'16 paper, they claims " all cameras are shifted towards a common focus plane while keeping the optical axes parallel" (Sec.4.1). Does it mean every image from this dataset have the same focus distance?

Also based on my observation, comparing to the demo focal stacked contained in this repo., there is no significant defocus on their images. May I ask how do you generate the focal stack from HCI 2016 dataset? I know in TIP paper, you mentioned it is related to Sec. 3-B.2, but I cannot understand what the relationship is.

Plus, I wonder if you know any other public available DFF dataset? I am trying to train a deep learning model for DFF tasks. But the only 2 datasets I found contain per image focus distance with ground truth depth are DDFF-12 and FoD, which in total only contains 1100 training samples.

Thank you!

jaeheungs commented 3 years ago

@fuy34

Sorry for the late reply, I must have missed the notification email.

  1. Yes I believe they are of the same focus.
  2. The focal stack is generated by using the GT depth data and applying blur based on that depth data.
  3. I'm not really aware of any public DFF datasets, so I'm sorry but I won't be of much help :S