CVPR2020-Camera On-boarding for Person Re-identification using Hypothesis Transfer Learning---论文阅读笔记

Abstract

他们这个是在to explore the dynamic nature of a camera network, where one tries to adapt the existing re-identification models after on-boarding new cameras(什么是on-boarding new cameras) 意思加入新相机后，还是尝试利用之前存在的re-ID模型。因为最近有a few方法that attempt to address this problem by assuming the labeled data in the existing network is still available (假设在existing network里的labelled data还可用什么意思？ 网络里会储存标签数据？) 当adding new cameras. 什么意思？ 不是数据的标签里还有包括是哪个摄像机组，组里哪个相机， 哪个identity, 这个identity的哪张么？ 那这个相机都没出现过，肯定没标签啊。 而且，因为隐私的限制，有人接触不到本来的数据，所以他们想到用存储在the learned re-identifications models, which mitigates any data privacy concern. 这就是他们说的他们develop an efficient model adaptation (利用模型的方法) approach using hypothesis transfer learning (假设迁移学习) that aims to transfer the knowledge using only source models (i.e., learned metrics) and limited labeled data(这个limited labeled data什么意思？) but without using any original source camera data from the existing network. 他们的方法minimize the effect of negative transfer by finding an optimal weighted combination of 多个source models for 迁移知识(意思，找到多个source models的最优加权组合，然后最小化negative transfer的影响)。

这篇文献解决了什么问题？

most of these works have not yet considered the dynamic nature of a camera network, where new cameras can be introduced at any time to cover a certain related area that is not well-covered by the existing network of cameras. 然后呢，Given newly introduced cameras, the traditional re-id methods aim to re-learn the pairwise matching metrics using a costly train phase. In this case, we can not afford to wait a long time to obtain significant amout of labelled data for learning pairwise metrics. Thus, we only have limited labeled data of persons that appear in the entire camera network after addition of the new camera.

1. How can we swiftly on-board new camera in an existing re-id framework without having access to the source camera data that the original network was trained on
2. How can we swiftly on-board new camera in an existing re-id framework relying upon only a small amount of labelled data during the transient短暂的 phase. 也就是：只依靠加入新相机后短暂时间内少量获得的数据。

忽视相机网络的动态本质，也就是随时可能introduce新相机来cover之前相机网络没好好cover的区域。然后呢，给了新Introduced的相机，传统的re-id方法就会去re-learn the pairwise mathcing metrics(成对成对地聚类.) 但是这太耗时了. 而且也没时间去等着说这个相机再收集了数据，然后把这些收集的带新相机label的数据加进去重新训练，说的有限的意思，就是在加入了新相机后，在整个camera network中出现的人只有limited labeled data

因为，本来应该所有相机的label都应该有，但是这样的话新加的没有，所以用Limited这个词！

这篇文献的创新点在哪？

• Only a few labeled identities that are seen by the target camera (新相机获得的数据也不是一点都没带label也带，自己新收集的有少量带，之前收集的被别的相机labeled的identities他这边能看到少量)， 此外，最小化the risk of negative transfer并且该算法的表现和fully supervised case接近。
• one or more of the source cameras, are needed for effective transfer of source knowledge to the newly introduced target cameras.(什么意思，一个或者多个source camera对于迁移source knowledge to 新相机是必须的)
• unlike [25, 26]的方法，which identify only one best source camera that aligns maximally with target camera(找出源camera中与目标camera中对齐maximally的那个)， 他这边是identify an optimal weighted combination of multiple source models(是source models还是source cameras)
• develop an efficient convex optimization formulation based on hypothesis transfer learning (多metric假设迁移学习算法) that minimizes the effect of negative transfer from any outlier (异常的) source metric.
• learn the weights of different source metrics and optimal matching metric jointly by alternating minimization 学习不同source metrics的权重，并最优的匹配metric. where the weighted source metric 被用作偏置正则项that aids in learning the optimal target metric only using limited labeled data.

• 第一个黄颜色的图没啥创新，就是告诉你度量学习咋聚类的
• 右上角的绿色图像，Provide pairwise metrics (提供成对的metrics意思是算好的数值比如$M_{12}$之间的distance metrics)
• 最下面的图，意思也给你加入新相机后的一些数据，这数据特点在于new camera with new limited pairwise labeled data(好奇怪，不是说不给access to source data么，那绿色，橘色，黑色框边的数据什么意思？)

什么是hypothesis Transfer Learning?

Hypothesis transfer learning is a type of transfer learning that uses only the learned classifiers from a source domain to efficiently learn a classifier in the target domain, which contains only limited labeled data.

这篇文献的路子：

把利用仅有的有限的标签数据(limited amount of labeled data across the target and different source cameras)还有pairwise source distance metrics，这些是应用场景前提，然后把这个问题描述为一个有约束的凸优化问题，然后这个数学公式aims to transfer knowledge from the source metrics to target. 这种方法不要求被new target 看过的identity在Source camera的存在性。请注意，我们的方法不需要在所有源摄像机中都出现在新目标摄像机中的每个人。

比如图中那种，ordered pair images across cameras target and source就是3对.

主要目标是to learn the optimal metric between target and each of the source cameras (找到target和每个source camera之间的最优的metric) by using the information from all the pairwise source metrics and limited labeled data.