search

gdlg / panoramic-depth-estimation

Eliminating the Blind Spot: Adapting 3D Object Detection and Monocular Depth Estimation to 360° Panoramic Imagery
https://gdlg.github.io/panoramic
Other
64 stars 10 forks source link

About training dataset & dataset transformation #6

Open yuniw18 opened 4 years ago

yuniw18 commented 4 years ago

Hi. I have questions about datasets used for training. In your paper, the best performance was achieved by using CARLA & Mapillary datasets. However, CARLA datasets which you uploaded, do not have stereo pair images. How the training proceeds using those datasets??

Also, I wonder how the rectilinear images were transformed to equi rectangular images. Any codes or reference links will be helpful. Thank you.

gdlg commented 4 years ago

Hi @yuniw18,

In our paper, we trained on the KITTI dataset, then evaluated on the CARLA dataset. This was the one main point of the paper: how to transfer trained weights from one dataset to another in the scenario where we don’t have enough information to train on the destination dataset.

The math for conversion between rectilinear & equirectangular is in our paper, however I can provide you with some Python code if you would like.

yuniw18 commented 4 years ago

Thank you for a quick reply. I misunderstood training datasets of your paper. Sorry for that.

And if you can give some python code, it will be very helpful!! Thank you

MasterHow commented 3 years ago

Hi @yuniw18,

In our paper, we trained on the KITTI dataset, then evaluated on the CARLA dataset. This was the one main point of the paper: how to transfer trained weights from one dataset to another in the scenario where we don’t have enough information to train on the destination dataset.

The math for conversion between rectilinear & equirectangular is in our paper, however I can provide you with some Python code if you would like.

Hi, I'm a new researcher trying to understand your great work! Shall I get some code for dataset transformation too? it would be very helpful! =)