pytorch_depth_from_videos_in_the_wild
pytorch_depth_from_videos_in_the_wild copied to clipboard
bolianchen
incorrect intrinsics results with provided pretrained model
Hello, thanks for your work on the code reconstruction. I got some problems when loading the model you provide in the readme page. I output the camera intrinsics and found there is a large distance with GT (taken in kitti raw). My output is as follow. I don't know where the problems lie in.
Hi Ljiaooo,
Do you mean the x and y offsets are large? I should say the learning of the camera intrinsics is not perfect. It is theoretically able to converge to the true matrix when data and training techniques are proper, and is convenient when you don't know the values.
Thanks for your reply, I found your reproduced IntrinsicsHead() is a little different from the original where the softpuls function is not applied on the offsets by the original. After I deleted the softplus on offsets, the camera matrix got normal.
@Ljiaooo Thank you very much for the findings. Since the learnings of depth and intrinsics are affecting each other, really appreciate if you may share some experiment results to show the improvement brought about by this revision and send a pull request.
Hello, Thank you for this repository, What is the image dimensions used for the intrinsic matrix above?
