PWC-Net
PWC-Net copied to clipboard
NVlabs
The "amount" of data augmentation
Dear Dr. @deqings : Thank you for your code.I have reimplemented the PWC-Net with tensorflow.In my code,I find that the net converge so slowly if augment on all of the training sets,say:FlyingChairs.The types of data augmentation are followed FlowNet2: 1.Translate all of images by a value between -20 and +20 percent of original size on x- and y-axis independently 2.Scale all of images by a factor between 90 and 200 percent of original size 3.No horizontal flip and vertically flip is used 4.Add the Gaussian noise that has a sigma uniformly sampled from [0, 0.04] 5.Add the contrast sampled within [−0.8, 0.4] 6.Multiplicative color changes to the RGB channels per image from [0.5, 2] 7.gamma values from [0.7, 1.5] and additive brightness changes using Gaussian with a sigma of 0.2. Should I not do data enhancements on all data？I would be very appreciated if you can help me:) Thanks and regards.
@guuusss when you remove the data augmentation, will the generalization ability of your network deteriorate?
My network will converge fast on FlyingChairs when I remove the data augmentation.
@ Thanks for your replying!! I find that when I remove the data augmentation, the network does converge better on Chairs but get the generalization ability worse on sintel.
@guuusss do you still working on flow? We have two guys here following PWC-Net. Would you be interested in joining our discussion group?
@xianshunw Sorry for my late reply, It is a great pleasure for me to join in your group. So, how can i take part in it?
@guuusss perfect!! we have a wechat group here, ~~IMAGE~~ If you don't use wechat, google talk and Skype are available for us.
A question: is Normalize (e.g. torchvision.transforms.Normalize in PyTorch) necessary in training optical flow estimator?
