search

sconlyshootery / FeatDepth

This is the offical codes for the methods described in the "Feature-metric Loss for Self-supervised Learning of Depth and Egomotion".
MIT License
247 stars 28 forks source link

Question about MS evaluation #75

Closed AttackonMuggle closed 2 years ago

AttackonMuggle commented 2 years ago

Thank you for the pre-trained model and test code. I just find that when evaluate the model trained by MS, if I use ratio like this, it will get better results than STEREO_SCALE_FACTOR. https://github.com/sconlyshootery/FeatDepth/blob/550420b3fb51a027549716b74c6fbce41651d3a5/scripts/eval_depth.py#L91 But in your config, you set stereo_scale = True if 's' in FRAME_IDS else False. Can you explain why you set it that way?

sconlyshootery commented 2 years ago

We do this for mono because the scale is ambiguous, we use this operation to find a best matched scale to compare with GT, but for stereo, baseline is known, still using this operation will be unfair to compare with other stereo methods.

sconlyshootery commented 2 years ago

Unsupervised Learning of Depth and Ego-Motion from Video this paper may help you get a better understanding of this setting.

AttackonMuggle commented 2 years ago

Thanks for your reply, and I have another question. Have you trained this code with only stereo data? My test results are not so good, but training on MS can bring improvement. I sincerely hope for your help.

sconlyshootery commented 2 years ago

Sorry, I didn't try on the Stereo data. This setting will not train posenet。

sconlyshootery commented 2 years ago

you can reopen it if you have further issues.