Keyword [Feature Generation]
Xian Y, Lorenz T, Schiele B, et al. Feature generating networks for zero-shot learning[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 5542-5551.
1. Overview
1.1. Motivation
- data imbalance between seen and unseen classes
In this paper
- synthesizes CNN features conditioned on class-level semantic information
- WGAN + classification loss
- exploit generated CNN features to train softmax classifier
1.2. Contribution
- f-CLSWGAN
- experiments on five dataset
- generate deep feature to improve ZSL
1.3. Related Works
1.3.1. GAN
- GAN
- cGAN
- DCGAN
- InfoGAN
- WGAN
1.3.2. ZSL and GZSL
- learn compatibility between images and classes
- classification
2. Methods
2.1. Definitions
- x. image feature
- y. seen label
- c(y). label embedding
- u. unseen label
2.2. f-GAN
- D. [0, 1] with MLP + sigmoid
2.3. f-WGAN
- α ~ U(0, 1)
- D. no sigmoid
2.4. f-CLSWGAN
2.5. Classification
Multimodal Embedding
Softmax
3. Experiments
3.1. Dataset
3.2. Details
- sentence. character-based CNN-RNN
- attribute.
- G and D. MLP + LeakyReLU
- no BN
- the final layer of G. ReLU
- λ=10, β=0.01
3.3. Evaluation
- T1. average per-class top-1 accuracy
- harmonic mean
3.4. Comparison
3.5. Ablation Study
