Imagenet
- PASCAL VOC: 20 classes, $3\times10^4$
- Imagenet-40
- ILSVRC (Imagenet-1K): 1000 classes, $1.28\times 10^6$
- IMAGENET: 20000 classes, $1.5\times10^7$
pretext tasks
- Crop
- Resize
- Flip
- Rotate
- Color distort
- drop
- jitter
- Gaussian noise
- Gaussian blur
- Sobel filtering
Contrastive Learning Framework
- 特点
- 在 feature space 上构建度量距离
- 通过特征不变性,得到多种预测结果
- 使用 Siamese Network
- 无需 pixel-level 重建
Generative Methods
- AE
- VAE
- GAN