FAQ for 4D Light Field Benchmark at lightfield-analysis.net
430 views
Skip to first unread message
Dierk Ole Johannsen
Sep 5, 2017, 10:52:40 AM9/5/17
to Light Field Vision
FAQ
[1] We are wondering what your camera setup in Blender is like? How come there are negative disparities if the cameras are parallel and identical?
Please look at our additional material for schematics. Technically speaking the cameras are not identical, but the sensors/image planes are shifted (not rotated!) towards the center of the light field. Otherwise, the optical axes and image planes for all cameras are parallel. The shift is equivalent to rendering a wider field of view and cutting out different parts for each view. Due to this the disparity range of a scene is not e.g. 30px to 34px, but -2px to 2px. I.e. in this example all views would be shifted be (S,T)*32px depending on their position in the light field. (S,T) corresponds to the difference (s,t)-(cs,ct), where (s,t) is the position in the light field and (cs,ct) the position of the center view.
The computation of the offset is described here: https://github.com/lightfield-analysis/evaluation-toolkit#7-compute-pixel-offset.
[2] How is the disparity map for the image that has to be uploaded generated? I mean, if you have 9x9 views, 81 images, you will end up with 81 disparity maps.
Comparing the center view to its neighbor yields the disparity map d. Comparing it to its neighbor in the other direction yields -d. The next neighbor would have two times the disparity (i.e. 2d).
In general, comparing a view (s,t) to the center view (cs,ct), yields disparity (S,T)*d, where (S,T)=(s,t)-(cs,ct). Thus, the disparity map is multiplied by the "baseline" between the center view and the other view. In this case, the "baseline" is without any units, i.e. e.g. just (1,7)-(5,5)=(-4,2).
[3] How do depth and disparity relate?
Due to the shift we use (see question [1]) the relation between disparity and depth is not simply z= focalLength*baseline/d but a little bit more complicated. The exact formulas are
β = baseline ∗ focalLength ∗ focusDistance ∗ resolution)
d = (β/Z − baseline ∗ focalLength ∗ resolution)/focusDistance/sensorSize
Z = β/(d ∗ focusDistance ∗ sensorSize + baseline ∗ focalLength ∗ resolution)
Please also see our additional material.
[4] I am confused about some parameters in the parameters.cfg file. What's the meaning of ...? Is there any reference for me to understand the above parameters?
Please check out or additional material. There is an in depth description of all parameters and schematics of the camera.
[5] Whats that fstop thingy?
Well... point taken, we did not write about this in the additional material. It's a blender internal parameter that simulates an aperture for the rendered camera. Usually, it should be set to high number (i.e. small aperture, usually 100 works well) to get an image which is not affected by depth of field.
[6] Can you send me the four test images' ground truth? I just want to get more data to test my method.
No ;) nice try. The ground truth for the test images is hidden to ensure a fair evaluation of all algorithms.
[7] I would like to use your data for a depth from focus project and for this reason, I need to generate focal stacks. I have gone through your dataset but I am a bit struggling to understand how easy it would be to generate the focal stacks.
You can find code to generate focal stacks from our data at www.lightfield-analysis.net/refocus.zip.
[8] Can we have access to your Blender files to e.g. render higher angular resolution?
We are currently working on sorting out copyright issues. We are able to share some of the scenes with open models. Please contact us directly for a link.
[9] I was wondering if you have the depth ground-truth data for all the views inside any scene. For now, the depth information is given only for the center view.
The depth of all views is only available for the additional scenes and can be found at the end of the email containing the download links.
[1] We are wondering what your camera setup in Blender is like? How come there are negative disparities if the cameras are parallel and identical?
Please look at our additional material for schematics. Technically speaking the cameras are not identical, but the sensors/image planes are shifted (not rotated!) towards the center of the light field. Otherwise, the optical axes and image planes for all cameras are parallel. The shift is equivalent to rendering a wider field of view and cutting out different parts for each view. Due to this the disparity range of a scene is not e.g. 30px to 34px, but -2px to 2px. I.e. in this example all views would be shifted be (S,T)*32px depending on their position in the light field. (S,T) corresponds to the difference (s,t)-(cs,ct), where (s,t) is the position in the light field and (cs,ct) the position of the center view.
The computation of the offset is described here: https://github.com/lightfield-analysis/evaluation-toolkit#7-compute-pixel-offset.
[2] How is the disparity map for the image that has to be uploaded generated? I mean, if you have 9x9 views, 81 images, you will end up with 81 disparity maps.
Comparing the center view to its neighbor yields the disparity map d. Comparing it to its neighbor in the other direction yields -d. The next neighbor would have two times the disparity (i.e. 2d).
In general, comparing a view (s,t) to the center view (cs,ct), yields disparity (S,T)*d, where (S,T)=(s,t)-(cs,ct). Thus, the disparity map is multiplied by the "baseline" between the center view and the other view. In this case, the "baseline" is without any units, i.e. e.g. just (1,7)-(5,5)=(-4,2).
[3] How do depth and disparity relate?
Due to the shift we use (see question [1]) the relation between disparity and depth is not simply z= focalLength*baseline/d but a little bit more complicated. The exact formulas are
β = baseline ∗ focalLength ∗ focusDistance ∗ resolution)
d = (β/Z − baseline ∗ focalLength ∗ resolution)/focusDistance/sensorSize
Z = β/(d ∗ focusDistance ∗ sensorSize + baseline ∗ focalLength ∗ resolution)
Please also see our additional material.
[4] I am confused about some parameters in the parameters.cfg file. What's the meaning of ...? Is there any reference for me to understand the above parameters?
Please check out or additional material. There is an in depth description of all parameters and schematics of the camera.
[5] Whats that fstop thingy?
Well... point taken, we did not write about this in the additional material. It's a blender internal parameter that simulates an aperture for the rendered camera. Usually, it should be set to high number (i.e. small aperture, usually 100 works well) to get an image which is not affected by depth of field.
[6] Can you send me the four test images' ground truth? I just want to get more data to test my method.
No ;) nice try. The ground truth for the test images is hidden to ensure a fair evaluation of all algorithms.
[7] I would like to use your data for a depth from focus project and for this reason, I need to generate focal stacks. I have gone through your dataset but I am a bit struggling to understand how easy it would be to generate the focal stacks.
You can find code to generate focal stacks from our data at www.lightfield-analysis.net/refocus.zip.
[8] Can we have access to your Blender files to e.g. render higher angular resolution?
We are currently working on sorting out copyright issues. We are able to share some of the scenes with open models. Please contact us directly for a link.
[9] I was wondering if you have the depth ground-truth data for all the views inside any scene. For now, the depth information is given only for the center view.
The depth of all views is only available for the additional scenes and can be found at the end of the email containing the download links.
Reply all
Reply to author
Forward
0 new messages