一、MCG-ICT-CAS Object Detection
1、获得成绩
![](https://img.haomeiwen.com/i7394071/cda6ec7e21274b47.png)
2、overview
![](https://img.haomeiwen.com/i7394071/7a3733423e3d6f27.png)
3、主要的创新点
(1)在建立tubelet时,提出了基于detection的检测方式
a.Adjacent Checking–By optical-flow for precise tubelet
If IoU (Red, Green) > a given threshold: same tubelet
else: a new tubelet
b.Detection Box Sequentialization
c.Coherent reclassification–Use majority voting to get coherent categories
throughout a given tubelet
![](https://img.haomeiwen.com/i7394071/ef09c5a01ca742a4.png)
但是发现效果不够好,会出现目标消失和轨迹离散的情况,于是想到将基于检测的tubelet与基于track的tubelet相结合
tracking based tubelet:(作者发现会出现位置漂移)
a.Anchor Frame Selection–Select the frame with highest detection score object(其他方法大部分都是用的这个方法)
b.Anchor Target Selection–Exploit the adjacent information with optical flow [5]to determine the reliable anchor targets
![](https://img.haomeiwen.com/i7394071/d4a1925fb9f9576a.png)
c.Anchor Target Selection: remove the unreliable去掉一些不可相信的检测框(通过bbox match)
d.Multi-target tracking with detection recall
(2)tubelet fusion:
有两种方式:
Union Fusion——Merge overlapping tubelets(有重叠部分)
![](https://img.haomeiwen.com/i7394071/0182bbcbd1a5afc6.png)
Concatenation Fusion——Merge successive tubelets(连接的)
![](https://img.haomeiwen.com/i7394071/210ac51612ff57df.png)
4、其他结构配置
still image detection:
![](https://img.haomeiwen.com/i7394071/f1ed4525862752c3.png)
网友评论