-
Notifications
You must be signed in to change notification settings - Fork 1
/
README.txt
62 lines (44 loc) · 2.71 KB
/
README.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
# JMMAC
This is an offical MATLAB implementation of tracker 'Joint Modeling Motion and Appearance Cues~(JMMAC) for Robust RGB-T Tracking' submitted to
VOT-RGBT2019 challenge.
## Introduction
In this work, we have found that both motion and appearance cues are important for designing a robust RGB-T tracker. The motion cue includes two
components: camera motion and object motion. The camera motion is inferred based on the key-point-based image registration technique; and the
object motion is estimated based on the camera motion estimation and the Kalman filter method. The appearance cue is captured based on an improved
ECO~\cite{ECO} model, where some complementary features (including deep and hand-crafted features) are selected for the RGB-T tracking task.
When the object suffers from heavy or full occlusion, a motion-guided tracking mechanism is used to avoid drifting, which makes the tracker be
dynamically switched between the tracking and prediction states. Our final tracker achieves 0.4826 of Expected Average Overlap~(EAO) in the
VOT-RGBT2019 challenge.
## Dependence
CUDA 9.0
CuDNN 7.1.0
Matconvnet 1.0-beta25~(www.vlfeat.org/matconvnet/download/matconvnet-1.0-beta25.tar.gz)
mcnSSD~(https://github.com/albanie/mcnSSD)
MATLAB 2015b~(Other version of MATLAB may get different results.)
gcc 4.8
g++ 4.8
PDollar Toolbox~(https://github.com/pdollar/toolbox)
## Operation System
Our tracker is run on Ubuntu 16.04 LTS with Intel i7-4790 @3.6GHz and Nvidia GTX TiTanX GPU. Other OS and GPU may get different results.
## Installation
Start Matlab and navigate to the repository.
Run the install script:
|>> install
if you have no GPU, you can run our tracker with single CPU.
Run the install_CPU script:
|>> install_CPU
## Integration Into VOT
Since MATLAB needs compilation and there will be a one-time delay for GPU computing commands, please increase the time limitation of trax by
setting 'trax_timeout' = 3000 in file 'workspace_load.m'.
To integrate the tracker into the Visual Object Tracking (VOT) challenge toolkit, check the 'VOT_integration' folder.
Copy the configuration file 'tracker_JMMAC.m' to your VOT workspace and set the path to the tracker reposetory inside it.
## Testing without GPU
If you run the tracker without GPU, please copy the configuration file 'tracker_JMMAC_CPU.m' to VOT workspace.
Note that this may lead a performace fluctuation without using GPU.
## Reference
@InProceedings{ECO,
Title = {ECO: Efficient Convolution Operators for Tracking},
Author = {Danelljan, Martin and Bhat, Goutam and Shahbaz Khan, Fahad and Felsberg, Michael},
Booktitle = {CVPR},
Year = {2017}
}