Question about pixel interval setting about cvp-mvsnet HOT 3 CLOSED

Thank you for pointing us to this paper.

Regarding your questions:
I would comment that the question should be more precise. 0.5 pix distance on 2D could mean a large interval on the visual ray if the centre point is far away from the camera or a very small range if the 3D point is close to the camera. It is also related to the camera parameters. I would answer your questions that it depends...

It could be easily figured out based on some simple geometric derivation. We actually should put it online. Stay tuned.

from cvp-mvsnet.

Hi Kwea123,

Thank you for your interest in our project.

As mentioned by @MiaomiaoLiu, the depth interval depends on multiple things including image resolution, intrinsic & extrinsic parameters, pixel location on image, depth of the pixel, etc.
Thus it really depends.
You can refer to Figure 3 of the newest version of our paper on arXiv for more information.

However regarding your concern, I can provide you one example of the calculated interval.
For the first view of scan1 of the DTU dataset, during evaluation on 1600x1200 image using 5 levels, the calculated interval for 4 refinement levels of our model is:

So the final depth interval is around 1.14mm for this view.
For other views it may vary based on above mentioned reasons.
Hope this address your concern.

Cheers,
Jiayu

from cvp-mvsnet.

Based on my own experiments (not the implementation of your paper) I found that the performance first increases as the interval shrinks, then it starts to decrease when the interval becomes smaller than a threshold. I started from 2mm (uniform for all pixels) and gradually decrease it; the performance is the best at about 1.5mm then it starts to go worse.

I wanted to know if the optimal interval of yours aligns with mine, so I asked for the interval in mm. It turns out that we have roughly the same result: As the interval decreases we first get increasing performance, but there is a optimal value beyond which the performance will start to decrease (Table 4b). This threshold is about 0.5pix in your findings, which translates to about 1.~ mm.

Finally I'm still left with the question that how can this paper reach 0.8mm without performance drop? Will wait for them to opensource..

from cvp-mvsnet.