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
61 stars 9 forks source link

Indoor datasets #8

Closed efyphil closed 3 years ago

efyphil commented 3 years ago

Thanks for your work! I tried to use your wights for predict indoor scenes, but get very bad result. Did you try to work with indoor?

Снимок экрана 2020-10-07 в 16 10 58
gdlg commented 3 years ago

Hi @efyphil , I worked on autonomous driving datasets and I haven’t tried indoor scenes. The algorithm was trained on the KITTI dataset. I think that the data that I trained on is too different from indoor imagery for the network to generalise well. You might need to train the algorithm on a dataset that contains indoor scenes.

efyphil commented 3 years ago

@gdlg thanks for fast answer. I'll try to train on different dataset, about bad result it's my fault. I didn't crop the image, after cropping it seems better

Снимок экрана 2020-10-08 в 11 57 38
gdlg commented 3 years ago

@efyphil Thanks for the update. For the images that I was working for, cropping was fine. It’s a limitation of the approach that it doesn’t work with the full vertical field of view.