search

zhyever / LiteDepth

Official Implementation of "LiteDepth: Digging into Fast and Accurate Depth Estimation on Mobile Devices"
54 stars 6 forks source link

About the number of upsample channels #2

Open puyiwen opened 2 years ago

puyiwen commented 2 years ago

Hi,a great work! I have some questions about the model. How do you set the number of upsampled channels? Whether it is derived by experiment or by experience? Thank you very much!!

zhyever commented 1 year ago

Both needed. I set channel numbers decreasingly (like 512, 256, 128, 64) by experience and reduce the numbers by experience (reduce each to a small number).

puyiwen commented 1 year ago

Thank you for your answer. Regarding the data enhancement part, I have always been very confused. I wonder if the depth map rotation doesn't change the ground truth? Do I need to supplement the depth value after cropping the image? For example, if the original image is 640×480, cropped to 600×400, and then resized to 640x480 (network input size), then it is equivalent to the camera position becoming closer, and the depth value should be scaled down?

puyiwen commented 1 year ago

When I was doing the experiment, I found a problem that your LiteDepth structure performs better than fastdepth on any indicator of NYU_Depth_v2 data, but when I took a few real indoor pictures alone for demo, I found that LiteDepth depth estimation effect is worse than fastdepth, why is this? Is my demo image accidental?

zhyever commented 1 year ago

After cropping, there is no need to resize it back to the raw resolution. The resizing will cause a change in pixel size, leading to performance degradation of models.

That's related to the model generalization. May the LiteDepth suffers from a generalization issue. That's also an interesting topic but there is little literature. I also observe this issue in my DepthFormer and BinsFormer. While the metrics are good, it is not practical for application. So, I think that is actually something important.