The use of DIML dataset

[LiheYoung]

The use of DIML dataset

[YvanYin]

Yes, this is a problem. Apart from that, some scenes' stereo images are not well rectified.

[YvanYin]

We only used the outdoor part of DIML.

[LiheYoung]

Thank you for your quick reply. However, according to the table (from DIML paper) below, there are 322K outdoor training images in DIML. And are there any reasons for discarding the indoor part of DIML, and only using the synthetic Taskonomy dataset for indoor training?

[JUGGHM]

Thank you for your quick reply. However, according to the table (from DIML paper) below, there are 322K outdoor training images in DIML. And are there any reasons for discarding the indoor part of DIML, and only using the synthetic Taskonomy dataset for indoor training?

We found some problems when pre-processing the data and did not use it.

[YvanYin]

Thank you for your quick reply. However, according to the table (from DIML paper) below, there are 322K outdoor training images in DIML. And are there any reasons for discarding the indoor part of DIML, and only using the synthetic Taskonomy dataset for indoor training?

1. The stereo calibration is broken in provided some data, thus we remove them. You could download them and check it.
2. DIML does not provide the camera intrinsics for the indoor part. We cannot use them in training.

[LiheYoung]

Thank you for your answer! About the first point, I find that the zip file of some scenes cannot be unzipped. Do you mean this by "broken"?

[LiheYoung]

Thank you a lot!