generalized mean pooling 参数generalized mean pooling参数 广义均值池化(GeneralizedMeanPooling)是一种常用的特征池化方法,它可以采用不同的参数$p$来控制池化的方式,从而更好地适应不同的数据特征。通常情况下,当$p=1$时,广义均值池化等价于普通的平均池化。当$p rightarrow infty$时,广义均值池化等价于最大池化...
在得出最终GCD结果时,其对应文本也会起到一个补充视角的作用,具体地,将k个检索语句的feature进行mean pooling以得到图像特征的另一个view。两个view结合之后可以直接在对应latent上做semi-kmeans即可。 实验 效果感觉还是不错的,不过CLIP在generic数据集上zero-shot的性能都不差,也符合情理。 下表中Knowledge代表是否...
Gu Y, Li C, Xie J (2018) “Attention-aware generalized mean pooling for image retrieval,” arXiv preprintarXiv:1811.00202 X.Wu, G.Irie, K.Hiramatsu, and K.Kashino, “Weighted generalized mean pooling for deep image retrieval.” In:2018 25th IEEE International Conference on Image Processing...
The constraint that the learned feature have zero mean is necessary to make the problem have a unique solution; otherwise we could add a constant to all feature values and obtain a different solution with equal value of the slowness objective.The constraint that the features have unit variance i...