search

fangchangma / sparse-to-dense

ICRA 2018 "Sparse-to-Dense: Depth Prediction from Sparse Depth Samples and a Single Image" (Torch Implementation)
Other
417 stars 95 forks source link

Question regarding dataset resolution #13

Closed ShreyasSkandan closed 6 years ago

ShreyasSkandan commented 6 years ago

If I read correctly, you use the KITTI odometry dataset in training this network.

I noticed that some image sequences in your hdf5 folder are of different dimensions than others. Pardon my ignorance if this is identical to the original dataset. If not, is there a reason for resizing them?

For example,

Sequence 03 is 1242x375, Sequence 00 is 1241x376, Sequence 04 is 1226x370 etc.

Thanks in advance, sorry if the answer to this is an obvious one.

Best regards, Shreyas

fangchangma commented 6 years ago

Hi Shreyas. The difference in sizes comes from the original dataset. In particular, I used the rectified images, and

The size of the images after rectification depends on the calibration parameters

see Section III.A (a) in the KITTI paper Vision meets Robotics: The KITTI Dataset

ShreyasSkandan commented 6 years ago

Great, thank you very much for the quick response. That makes complete sense now.