Performance of FlowNet2-CSS-ft-sd

[yxqlwl]

After testing FlowNet2-CSS-ft-sd on sintel training dataset, I found that:

1. On sintel clean, FlowNet2-CSS-ft-sd can produce exact the same AEE (2.08) as reported in <FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks>;

2. However, AEE on sintel final is 3.57, which is higher than the result reported in the paper.

So could you please check the performance of FlowNet2-CSS-ft-sd on sintel final? Thanks~

[brjeon]

Hi yxqlwl,
Have you performed training also with the provided prototxt?

[yxqlwl]

Hi,
No I didn't. I only used the provided deploy.prototxt and caffemodel.

[yxqlwl]

Has anyone also tested performance of the models on sintel final? Haven't found the bugs in my testing code yet.

[nikolausmayer]

Hi,
We will look into this.

Best,
Nikolaus

[eddy-ilg]

Thanks for the comment, will check it this week.

[eddy-ilg]

We confirm this issue. We originally made a dataset for sceneflow, where we took the sintel flow and disparity datasets together. The CLEAN images of both datasets are exactly the same, however the disparity FINAL images turn out to come without the atmospheric effects. I.e. we tested on Sintel with motion blur but missing the effects like fog.

We apologize for the inconvinience and will fix the numbers in the paper.

[yxqlwl]

Hi,
That's OK. We hope the performance of FlowNet2 could be further improved.
Best,
ViJay