Multimodal Contrastive Training for Visual Representation Learning

parameterize the image encoder as f${}_{iq}$

query feature q${}_{ii}$，key feature k${}_{ii}$
parameterize the textual encoder as $f_{cq}(\cdot ;{\Theta }_q,{\Phi }_{cq})$，momentum textual encoder as $f_{ck}(\cdot ;{\Theta }_k,{\Phi }_{ik})$.$c_j^{†}$和$c_j^{\star }$是different augmented examples

吐槽

第一张图字母下标被黑色背景盖住了，且作者不公布代码，不该是CVPR的“水平”