search

ucbdrive / hd3

Code for Hierarchical Discrete Distribution Decomposition for Match Density Estimation (CVPR 2019)
BSD 3-Clause "New" or "Revised" License
204 stars 31 forks source link

Large EPE error for stereo task #32

Closed mli0603 closed 4 years ago

mli0603 commented 4 years ago

Hi,

Thanks for your code! When I run the inference code to evaluate disparity estimation on sceneflow dataset, the EPE is very large (it is 121 from a test image) from line 42 (https://github.com/ucbdrive/hd3/blob/master/hd3model.py#L42). I visualized the output of the network curr_vect, I see all estimation are negative (https://github.com/ucbdrive/hd3/blob/master/inference.py#L197). I can confirm I am using the stereo network since I am using the model hd3sc_things-57947496.pth with context mode on.

Is the EPE reported using pixel? Is this normal? I am not sure if the EPE is only for optical flow or can be used for disparity estimation as well.

Thank you very much!

mli0603 commented 4 years ago

It seems like during EPE calculation, the output is negative disparity, while the ground truth is positive. I inverted the sign the error becomes much smaller.

Is this a problem of data or this is a mistake in EPE calculation?

yzcjtr commented 4 years ago

Hi, the EPE metric shouldn't be that large. The implementation should support both optical flow and stereo matching tasks. I guess you are using the original scene flow dataset instead of its subset. Please refer to a related discussion here (https://github.com/ucbdrive/hd3/issues/27).

mli0603 commented 4 years ago

@yzcjtr Oh yes, this is exactly the problem. Thank you very much! I appreciate it.