Can I give the model the GT pose instead of the model's pose network?

[sjg02122]

Thank you for the excellent work.

I'm doing an experiment on Depth Network. To this end, I would like to add GT of Pose instead of Posenetwork. How can I handle it?

Thanks!

[VitorGuizilini-TRI]

Sure, you would have to change this line here, so instead of using the predicted poses you pass the GT poses from the batch directly.

https://github.com/TRI-ML/packnet-sfm/blob/c03e4bf929f202ff67819340135c53778d36047f/packnet_sfm/models/SfmModel.py#L141

[sjg02122]

Thanks for your reply.

I just changed the code

https://github.com/TRI-ML/packnet-sfm/blob/c03e4bf929f202ff67819340135c53778d36047f/packnet_sfm/datasets/kitti_dataset.py#L97

with_pose --> True

then, I check the form ob data

the computed pose is list (length 2, [object of pose, object of pose]) gt pose in the batch (batch['pose'] is tensor, batch['pose_context'] is list(length :2 ,tensor))

how to change the code??

Thanks a lot!

[VitorGuizilini-TRI]

You will have to change the internal code, there is no flag to switch from predicted to GT pose, unfortunately.

[hjxwhy]

Thanks for your reply.

I just changed the code

https://github.com/TRI-ML/packnet-sfm/blob/c03e4bf929f202ff67819340135c53778d36047f/packnet_sfm/datasets/kitti_dataset.py#L97

with_pose --> True

then, I check the form ob data

the computed pose is list (length 2, [object of pose, object of pose]) gt pose in the batch (batch['pose'] is tensor, batch['pose_context'] is list(length :2 ,tensor))

how to change the code??

Thanks a lot!

Hello, have you changed success? I used the gt_pose, but the predicted depth is wrong. If you have changed success, can you show me where and how are you change the code?

[sjg02122]

        pose = None
        if 'rgb_context' in batch and self.pose_net is not None:
            pose = self.compute_poses(batch['rgb'],batch['rgb_context'])

Change the value of pose to GT. you can acess the gt value in batch dict

file : SfmModel.py

[livey]

Thank you for the excellent work.

I'm doing an experiment on Depth Network. To this end, I would like to add GT of Pose instead of Posenetwork. How can I handle it?

Thanks!

Have you tried this? When I use the GT pose, the self-supervised model does not produce a good depth estimation.

[GANWANSHUI]

@livey Hi, may I ask which dataset you are using for the GT-pose experiment? Thanks a lot!

[livey]

@livey Hi, may I ask which dataset you are using for the GT-pose experiment? Thanks a lot!

I used the DDAD dataset.

[VitorGuizilini-TRI]

The KITTI dataset seems to have poor GT poses, we routinely get worse depth estimates using GT poses compared to using predicted poses (the predictions are scaled, though).

[GANWANSHUI]

@VitorGuizilini-TRI @livey Hi, Thanks a lot for the information. According to the discussion in https://github.com/TRI-ML/DDAD/issues/29, it seems the GT-pose in DDAD dataset is also not well. @VitorGuizilini-TRI was there a refined version for the GT-pose for DDAD dataset? Thanks again!

[livey]

@VitorGuizilini-TRI @livey Hi, Thanks a lot for the information. According to the discussion in https://github.com/TRI-ML/DDAD/issues/29, it seems the GT-pose in DDAD dataset is also not well. @VitorGuizilini-TRI was there a refined version for the GT-pose for DDAD dataset? Thanks again!

When using the ground truth pose, I think the real problem is the dynamic objects.

[GANWANSHUI]

@VitorGuizilini-TRI @livey Hi, Thanks a lot for the information. According to the discussion in https://github.com/TRI-ML/DDAD/issues/29, it seems the GT-pose in DDAD dataset is also not well. @VitorGuizilini-TRI was there a refined version for the GT-pose for DDAD dataset? Thanks again!

When using the ground truth pose, I think the real problem is the dynamic objects.

Thanks for the discussion. Could you give more explanation about the dynamic objects？ I try to visualize the transformation of two random frames, and it seems reasonable. Thanks again!