regarding inverse_warping about sfmlearner-pytorch

[ClementPinard]

Hello,

thank you for your interest in this repo.

Actually, the function inverse_warp does implement a 6DoF pose to extrinsics function :

SfmLearner-Pytorch/inverse_warp.py

Line 136 in 4e6b7e8

def pose_vec2mat(vec, rotation_mode='euler'):

So you would have to change the code a little to use matrices instead of pose vectors, and thgen this line will have to go :

SfmLearner-Pytorch/inverse_warp.py

Line 177 in 4e6b7e8

pose_mat = pose_vec2mat(pose, rotation_mode) # [B,3,4]

One quick warning with using matrices instead of pose vectors is the output of thge network. It's much more stable to use pose vectors than extrinsics matrices as output to train if every ceofficient of the outputted matrix is a learnable parameter. So the best strategy for you is to have the euler2mat function directly embedded in your network that you will call at each forward.

That way, your network will output directly matrices instead of pose vectors, allowing you to do the inverse wapr with matrices, and will still be stable as if it was using pose vectors.

Hope it helped,

Clément

[wtpro]

Thanks for your reply!
Yes I am aware that pose can be converted to a matrix by using pose_vec2mat function.
What I am asking is whether it is possible to find the equivalence of pose in the form of extrinsics and intrinsics

[ClementPinard]

The code already uses an intrinsics matrix separately from pose, so the pose is totally independent from intrinsics.
when we convert pose to a 4x3 matrix, we actually get the extrinsics matrix, so there they are equivalent.

[wtpro]

So which extrinsics is pose equivalent to? The target image extrinsics or the source image extrinsics? Or is it
extrinsics_src @ extrinsics_tgt.inverse()
Judging from the code, I believe pose represents the transformation from target pixel coord to source pixel coord, so a single extrinsics might not suffice.

[ClementPinard]

In this setup, the extresincs of target image is always the identity matrix (or the null pose vector)

There is no way for the network to know the pose of both target and reference image, it can only estimate the difference between the two. In other words, it can estimate the extrinsincs of reference image in a coordinate system centered around target image.

If you want the pose of both target and reference images, you need an anchor somewhere. If the anchor is e.g. the first image of the whole sequence, and you need the pose of the Nth image relative to the first frame, you will need to accumulate the pose differences and thus compute the compisition of several extrinsics because they are not written in the same coordinate system.

[wtpro]

Sorry for the unclear quesrtion. I am currently trying to accommodate your inverse_warp to my own dataset without using the pose prediction network, because my dataset provides extrinsics and intrinsics.

I technically could just use the network for pose and call it a day but I just want to explore more possibilities considering the extrinsics and intrinsics are easy to obtain with specialized hardwares so a network is not needed.

[ClementPinard]

Oh, now I understand, sorry !

If you have extrinsics for both view, then yes the inverse warp does use the difference of extrinsics inplicitly, which results in your proposed formula in your first pose.

Now the tricky part is to make sure the order is right. Is it intrinsics @ extrinsics_src @ extrinsics_tgt.inverse() @ intrinsics_inverse() @ target_depth or is it intrinsics @ extrinsics_tgt.inverse() @ extrinsics_src @ intrinsics_inverse() @ target_depth ? Honestly, the best way to be sure is trial and error by visualizing the result.

Hope it helped !

[wtpro]

I did some quick visualization and here are the results:

intrinsics @ extrinsics_src @ extrinsics_tgt.inverse() @ intrinsics_inverse() @ target_depth:

intrinsics @ extrinsics_tgt.inverse() @ extrinsics_src @ intrinsics_inverse() @ target_depth:

target_image:

it seems that the first one is more correct to me. The second one contains some pixels that was unseen in target image.

Appreciate your input here!

[ClementPinard]

Hard to tell without the depth (are you using groundtruth from the sensor ?) but I'd agree with you that first one looks better, especially for the chair onn the bottom.

Note that the duplication of pixels in the first is normal for view that are occluded on ref image but not on tgt image. It's impossible to reconstruct them since they are not visible and thus the algorithm take the color of foreground. It is visible for the table in bottom as well for example.

[wtpro]

This is the GT depth of the target image. It is actually taken from a synthetic dataset not real life scene.

Also I have a small question, if the source and target images do not have any overlap, does this warp still work?