之前看过台大李宏毅的SVM介绍，我寻思我没看到有关于model的东西啊。
有点印象就是SVM跟全连接神经网络的训练方式是有点像的。
但是有关于SVM模型介绍是有点少的。
最近看到一篇paper
The relation probabilities are determined by a probabilistic SVM trained on 3D motion indicator features, and are incorporated into the group model probability (eqn. 2)
然后论文给出了它的训练方式
总的来说这是在训练一个group模型三类events的回归模型 ，
先不管它的原文是训练什么模型，
关键是它的real world annotated data也不给我个reference 我也要啊。。
看了参考文献估计是同一个实验室的data
hahaha我猜到了这里就是他的创新，果然在conclusion看到了
Here, we extend the approach of [16, 2] by the datadriven probability for continuation events p Gi C . The likelihood for continuation scales inversely with the highest probability for a non-continuation of that group (if, for instance, p GiGj M = 1.0, then p Gi C must be 0.0). Without this term, continuation events are overly biased which causes the tracker to follow splits and merges with unnecessary delays.
学到了，没什么太大的创新性的时候就说exdend this approach
We extended this approach by new expressions for group model probabilities and a book keeping logic to maintain stable group track identifiers robust to sporadic identifier switches of the underlying person tracks. The experiments demonstrate the viability of the approach on a real-world, unscripted outdoor RGB-D dataset collected with a mobile platform in a busy urban pedestrian zone.
我寻思我既没有training data 也没有人家成熟的团队，我看这个干几把？
但是之前好像在ICRA 2018 best paper 是用LSTM 做的同样的事情
The factor graph representation lends itself naturally to the S-RNN architecture [10].
看来我记错了，在看看Ｓ－ＲＮＮ是个什么变种ＲＮＮ ？
原来这个best paper 还有个爹 social LSTM
还好social LSTM有training data
所以说这个ＬＳＴＭ每个人标配一个ＬＳＴＭ然后每个ＬＳＴＭ网络share weights
agnostic 应该是怀疑论者的意思吧，意思就是single person 's trajectory impacts the neighborhood
and then introduce a new pooling strategy
every LSTM is connected through a spatially proximal LSTMs
这一步应该有点难理解，论文接下来
假设有三个人在person i周围嚯，然后这个判断这个neighbor的相对位置
看了一下这个只提供数据集接口啊
https://www.zybuluo.com/ArrowLLL/note/981714
还有个老哥也做了笔记
但是我寻思，就没有把这些model deployment 的吗。。