independent-research

Home Independent Research Last Next

Index

Z. Ge, S. Demyanov, Z. Chen, and R. Garnavi, “Generative openmax for multi-class open set classification,” arXiv preprint arXiv:1707.07418, 2017.

Motivation

The paper proposed a new model for openset problem, which “extends OpenMax by emoloying generative adversarial networks (GANs) for novel categoriy image synthesis”. That is to say, the proposed method uses GANs genrating unknown labels then sends unknown labels to OpenMax as well as labeled ones, thus get an open set image classifier.

Approach

The method proposed has two stages as well as OpenMax: Pre-Network training and score calibration. During Pre-Network traning stage, different with openmax, it first generate some unknown class samples (synthetic samples) then sent them along with known samples into networks for training.

The generated unknown samples also score calibrations.

The difference is it used K+1 classes instead of K classes in both stages. And the 1 (unknown class) was generated beforehand, whose objective function as:

Being different from normal GANs,

The function has an extra variable: C, which denotes categories, which will be random selected as well: D(data, category).

Limitation