Loss Question

[wuaming]

When you compute the epe loss, do you normalize the optical flow?

[liruoteng]

no, the epe loss is just L2 norm.

[wuaming]

How long do you take to train the model?

[liruoteng]

If you train it on GTX 1080 level GPU card, it will probably take you at least 3 days to fully complete.

…

On 9 July 2017 at 22:44, wuaming ***@***.***> wrote: How long do you take to train the model? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#18 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHIGm-iIVJt-XxYtK_fhywZSYOP7khZ2ks5sMOc5gaJpZM4ORqBg> .

[wuaming]

The size of input image is not equal the output size of flow. But how do you process the groundtruth? And how do you resize the flow to your output size?

[wuaming]

When you train the model, dose the dropout affect the accuracy of the model?

[liruoteng]

Hi, for the image size problem, you may want to see the resample layer in the train.prototxt, or deploy.prototxt.
And there is no dropout in this model.

thanks.

[wuaming]

When I finish the train, the result is not smooth.

[wuaming]

When you visualize the flow, do you use the smooth operation?

[wuaming]

Dear Professor, where is the code about data augmentation?

[liruoteng]

Hi, regarding visualization, we don't use smooth operation.
For data augmentation, please refer to the FlowNet paper.

Thanks
RT

[JasonZhangFan]

I'm wondering if we want to use it for real application, how can we get ground truth of the input as training material?

[liruoteng]

Hi, the ground truth is always difficult to collect and calculate for real data. You may follow this paper to see how they generate KITTI dataset