search

noahzn / Lite-Mono

[CVPR2023] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation
MIT License
540 stars 61 forks source link

the results of training are black pictures #147

Closed Junoin closed 3 months ago

Junoin commented 3 months ago

Nice work！ I encountered a problem during training. Using the given parameters for training on the kitti dataset, the loss dropped from 0.14 to 0.11, but the depth map I got was a completely black picture or a few black blocks. Does anyone know how to solve this problem?

noahzn commented 3 months ago

Hi, how did you visualize your depth maps? Did you check it in the tensorboard?

Junoin commented 3 months ago

Hi, how did you visualize your depth maps? Did you check it in the tensorboard?

Yes. The disp_0 disp_1 disp_2 in the tensorboard are all black pictures. And I also ran the test_simple.py using the weights I trained, but the results are alse very terrible :(

noahzn commented 3 months ago

Were they black from the beginning or they suddenly became black during the training? Did you change anything in the code, for example the intrinsics?

Junoin commented 3 months ago

Were they black from the beginning or they suddenly became black during the training? Did you change anything in the code, for example the intrinsics?

Thank you very much for your tips. I checked the parameters and found that the input dimensions of the pre-trained model and the model did not match. After the modification, the displayed disp is the depth map. Thank you again for sharing!

noahzn commented 3 months ago

Good to know!!