# Tracking Subjects and Detecting Relationships in Crowded City Videos -- Paper Review

## Contribution
1. 解決物體被遮蔽、模糊、交錯等造成的ID switch tracking問題，因不需要使用AI方法提取外部特徵，速度上比較快
2. 偵測人與人之間的肢體關係

## Methods

***IOU matching(baseline)***
幀與幀之間，將IoU重疊度最高的兩個框分配為同一個物件id

***cache(propose)***
針對上一幀所偵測到的，若下一幀該物體未再出現，先暫存到cache中，之後再判斷是被遮蔽了，或者是已離開畫面

***re-id(propose)***

## Result

IDs = number of ID switching (lower better)
MOTA = (IDs + FP + FN -1 )/#ground truth annotation (higher better)

* comparison on MOTA17
* comparison on MOTA17
* comparison on MOTA20
* relationship recognition

## Comments

pros:
Provides pure algorithm methods: cache and reidentification to deal with subject occlusion issue.

cons:
1. In ***section 6.3 - Comparison with the State-of-the-Art***, the FPS of proposed method shown in Table3 has sharply decreased from 122 (dataset MOT17) to 8.0 (dataaset MOT20) is a concern.
2. In ***section 7 - Evaluation of Relationship Discovery***, the sample size used in evaluating the accuracy of recognizing couples and families (Table4) is too small.
3. It is recommended to have pseudo code or public source code reference in the paper.