Demon Output about demon HOT 10 CLOSED

I also have the same question. I got to work the system but I don't know what the values are. Meters? Inches? Are they normalized? Please throw some light on this matter!

from demon.

They are calculating inverse of depth i.,e 1/d in the depth map that is outputted. And I believe its in metres.

from demon.

Also note that the predicted inverse depth matches the unit translation.

from demon.

As @sampepose says, the depth matches the unit translation.
Therefore, the depth does not have a unit.

@Zuriich the following issue is related to your question: #17

from demon.

Okay, well now I was applying the inverse of depth and I got another question: which type of preprocessing you have applied to the data before the training? (normalization, max and min value of this, etc).

from demon.

I was seeing the evaluation.py file and there appear three different ways of normalization (abs, log and inv). Which have you use to obtain the model you released?

from demon.

In general it is not possible to infer absolute scale values from any number of images.
Because of this we trained our network to predict a unit translation vector.
This means the translation length is always 1.

The scale of the depth values correspond to this unit vector and thus don't have a unit.
To recover actual depth values in meters it is necessary to know the length of the translation vector in meters.

Assuming you know this length (camera baseline) B in meters you can compute the depth as

B * 1/D

with D as the inverse depth returned by the network.

@Zuriich During training we scale the depth values and the motion such that the camera baseline is 1. There is no other normalization.

@edcross The abs, log and inv you see in the evaluation.py are ways of optimally scaling the depth values. This was meant for a comparison with other methods that do not compute a camera motion but infer scale. In the end we only used the scale invariant metric, which does not need this.

from demon.

Hello @benjaminum. I am using this thread since my question is related. Let me know if you want me to open a new thread.

I am trying to make sense of the optical flow that is coming out of the iterative_net, and I have noticed that all the values are always between -1 and 1 (I am using your sculpture image example). This does not seem right to me. I was expecting something from -10 and 10 maybe. Also, from the two components in the result['predict_depth2'], which one is x (horizontal) and which one is y (vertical)? If x points to the right, does y point down?

from demon.

from demon.

I think all questions here have been answered.

Thanks to everyone!

from demon.