博文

CVPR2019 PaperReading(1)

图片
[ ] On the Structural Sensitivity of Deep Convolutional Networks to the Directions of Fourier Basis Functions[ ] Striking the Right Balance with Uncertainty[x] NM-Net: Mining Reliable Neighbors for Robust Feature Correspondences[x] C2AE: Class Conditioned Auto-Encoder for Open-set Recognition[ ] Doodle to Search: Practical Zero-Shot Sketch-based Image Retrieval[x] Zero-Shot Task Transfer[ ] C-MIL: Continuation Multiple Instance Learning for Weakly Supervised Object Detection[ ] Transferrable Prototypical Networks for Unsupervised Domain AdaptationZero-Shot Task Transfer 这篇文章是针对zero shot的任务提出了一个新的meta learning的算法——TTNet。
这个meta-learner 学习模型参数时从有ground truth的已知任务来迁移到新的zero-shot任务中来。
本篇paper是做的taskonomy的zero-shot任务,以下4个任务作为zero-shot: surface-normal room layoutdepthcamera pose estimation Method 定义已知ground truth的m个task为, zero-shot task为.
用多任务学习的方法,将需要学习的meta-learning function 建模为有m个branches的network,参数分别为,这些任务分支在开头没有联系,在后面通过一个的block来coupled。
因此,可以分为两个部分,第一个部分是m个,第二个部分是。 疑问点: 本篇paper的TTnet是生成data n…