search

simonmeister / UnFlow

UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss
MIT License
294 stars 57 forks source link

Evaluation results compared with the paper #52

Closed jytime closed 6 years ago

jytime commented 6 years ago

Hi Simon,

I am testing the pre-trained models on the Kitti. In your paper (https://arxiv.org/pdf/1711.07837.pdf), the best unsupervised AEE (All) of UnFlow-C on Kitti 2015 training set is 8.80. But when testing, it returns:

python eval_gui.py --variant=train_2015 --ex C

(C) EPE_noc = 3.185742938518524 (C) EPE_all = 4.673751163482666 (C) outliers_noc = 16.443058681488036 (C) outliers_all = 21.893028831481935

May I ask is this because you have improved the model? Thanks a lot.

simonmeister commented 6 years ago

In the eval_gui.py file, we have a default of

tf.app.flags.DEFINE_integer('num', 10, 'Number of examples to evaluate. Set to -1 to evaluate all.').

Thus the numbers you report only give the errors over the first 10 examples from about 200. If you run

python eval_gui.py --variant=train_2015 --ex C --num -1

it should be equal to the results in the paper.

jytime commented 6 years ago

Thank you Simon