[ Archived Post ] Jointly Optimize Data Augmentation and Network Training: Adversarial Data Augmentation in Human Pose Estimation

Please note that this post is for my own educational purpose.

Can we optimize data augmentation → while training a classifier → all together → the augmentation → makes the classifier stronger → and human pose estimation → achieves a much stronger result.

Data collection is hard → also costly → even if a large amount of data is collected → might have limitation → now what if we augment the data augmentation procedure?

Like above → we can streamline the whole process → scaling rotating and more → very interesting procedure. (so we are now optimizing classifier with the network). (human pose estimation → this is very cool → another vision task can also be applied). (so this is playing around with a generator). (augmentation network is carefully designed → sampled to create new data points).

Adversarial learning → GAN → this line of research is very closely related to → a very good and sexy approach. (hard example mining → also related → focus on hard examples). (human pose estimation → already very good approaches).

The authors approach → visualized. (random data augmentation is good → but does not follow training procedures → a good method is to transform the given image → very interesting) (the generator can generate harder examples)

There is a model that evaluates the generated samples → and the generator gets to be trained at the same time. (policy is used → seem to be related to RL). (create new data point → that does not even exist in the data set).

Human pose estimation → U-net is used → very good model for biological data as well → reward and penalty method is used as well. (wow quite an interesting method of augmentation → definitely create images that are realistic)

Occlusion is also made → this method is also good → and deep features are occluded. (the mask is scaled up). (adding these constraints is an important job). Joint training → augmented images → does not have a ground truth → here reward and penalty policy is given → a prediction of the augmentation network → must follow some policy. (compared to adversarial augmentation). (a very smart way of combining DNN with RL).

How the network changed over time → while training. (popular hourglass is the target network → cool → residual module is used).

The authors' method → mostly gives the state of the art result for different human body parts.

Very good result → this kind of augmentation is powerful → and generates images in a realistic setting.

A new method of data augmentation → involving policy as well.

https://jaedukseo.me I love to make my own notes my guy, let's get LIT with KNOWLEDGE in my GARAGE