search

noahzn / Lite-Mono

[CVPR2023] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation
MIT License
527 stars 58 forks source link

Output resolution problem #151

Closed swing2331 closed 4 weeks ago

swing2331 commented 4 weeks ago

Hi I tried to train with a custom dataset of 640480 and tested it on the .pth I got using test_simple.py. Why is the resulting .npy depth map size 640192 instead of the 640*480 I set?

swing2331 commented 4 weeks ago

My data set is aerial images taken by drones, and the test results are as follows: Original image，0000001544 Depth Map，0000001544_disp .npy Depth Map，sd1

noahzn commented 4 weeks ago

Hi, in the options you can change the width and height. Did you change the camera intrinsics before training the network?

swing2331 commented 4 weeks ago

Yes, I changed the size in the opening, but the. npy size output by test_Simple.py is still 640 * 192

noahzn commented 4 weeks ago

Can you validate that the size you used for training is correct? Because the image size is read from the trained model, see the code here. I guess you trained your model using 640*192

swing2331 commented 4 weeks ago

Yes, upon inspection, it was found that I used a size of 192 * 640 for training. I will retrain. Thank you for your answer!