已经有两个人用opentld写论文了
gnebehay的程序,经测试不太满意,准确率有问题
zhang的程序,对熊猫的跟踪各有优势,但是对jumpinger的效果好,对david的跟踪有偏移。
其实opentld当track失败时,要降低detector的符合要求的阈值,这能减少一些失去跟踪的情况,这可能会有跟踪错误的情况,但至少不会有屏幕上全部空白的情况吧。还有tld的各项参数应该可以根据不同的视频类型进行调整吧。
还有tld应用于face检测的效果很不错,而且可以识别人脸,怎么做到的呢
tld的matlab版
Control keys:
=========================================================
n
... shows negative examples in online model (default on)
p
... shows positive examples in online model (default on)
i
... initialization of different target
c
... show confidence score (default on)
o
... show output as circle/dot/no output (default circle)
d
... show detections (default on)
t
... show target in top left corner (default off)
r
... replace target with first patch (default off)
#
... draw trajectory of target (default off)
<space>
... svae current image
1
... mode without learning (fast, does not improve)
2
... mode with learning (slower, but can learn new appearances and discriminate)
q
... finish application
.
... show points that are used for tracking
h
... show help
=========================================================
可以参照,另外加上cuda的技术