Compile | CV
Report | I love computer vision (wechat ID: aicvml)
Today, openmmlab, an open source organization of the multimedia laboratory (mmlab) of the Chinese University of Hong Kong, released a new member: mmtracking, which aims to promote research and development in the field of target tracking.
Many open source libraries under openmmlab, such as mmcv, mmdetection and mmaction2, have many users and become important tools for reproducing top conference papers and realizing project applications.
The openmmlab open source mmtracking Covered directions includeSingle object tracking, multi-target tracking, video object detection (VID), which not only includes the implementation of cutting-edge algorithm, but also has a mature pre training model for direct use.
Officials say this is the first unified video perception platform, which puts single target, multi-target tracking and video target detection in one framework;
Modular design divides the video perception framework into different modules to facilitate algorithm development;
- Simple, fast and robust.
- Simple: it is extremely simple to interact with other open source libraries of openmmlab. In particular, mmtracking and mmdetection are natural. You can switch detectors by modifying the configuration file.
- Fast: all operations are carried out on the GPU to ensure that the program runs fast, and even the implemented algorithm is faster than the implementation, training and testing of other open source libraries.
- Robust: the author has implemented a large number of state of the art methods, many of which are even better than the official implementation.
Currently implemented algorithms:
Video target detection:
DFF （CVPR 2017）
FGFA （ICCV 2017）
SELSA (ICCV 2019)
Multi target tracking:
SORT/DeepSORT （ICIP 2016,ICIP 2017）
Tracktor （ICCV 2019）
Single target tracking:
SiameseRPN++ (CVPR 2019)
Thanks to the high-quality implementation of the openmmlab open source framework, the official said:
Compared with the official implementation, the implementation result of video target detection algorithm selsa is on Imagenet vid dataset [email protected] 1.25 points exceeded.
Compared with the official implementation of the multi-target tracking algorithm tracktor, Mota exceeds 4.9 points and idf1 exceeds 3.3 points on the mot17 data set.
Compared with the official implementation of the single target tracking algorithm siameserpn + +, norm precision exceeds 1.0 points on the lasot dataset.
Although there are not many algorithms implemented at present, in view of mmlab’s strong R & D strength and community appeal, I believe that mmtracking is an Open-Source Library worthy of continuous tracking for friends in the tracking direction.
reference resources: https://zhuanlan.zhihu.com/p/…