In eco/params.h/search_area_scale = 4.5, if change to 2.5, ECO will lose the target in UAV123/person23.
While, in kcf/kcftracker.cpp/padding = 2.5, if change to 4.5, KCF will follow the target in the same data above.
In eco/params.h/nSamples = 50, if change to 10, ECO will lose the target in VOT/girl.
But, in eco/params.h/nSamples = 50, if change to 10, ECO will keep the target in UAV123/bike1.
The reason is for VOT/girl, it needs more history memory to return back to target. But for UAV123/bike1, long histroy means cannot adapte tothe recent changes. There is a dilemma in here.
For the localization, eco/feature_extractor.cpp/get_hog() occupies most of the time, so diminute the eco/params.h/number_of_scales will speed-up the algorithm. For gpu vesion, should try to rewrite this function.
For the training, if just sampling update, the time it takes is quite few. But if training (every eco/params.h/tarin_gap time), it takes lot of time.
Dataset that ECO_HOG get lost, KCF keeps correct: UAV123/wakeboard1
Dataset that ECO_HOG get lost, GOTURUN keeps correct: VOT/iceskater1 TLP/IceSkating
Dataset that ECO_HOG get lost, ALL the others keeps correct: TLP/Drone3
Just ECO_HOG right: TLP/Sam
No one right: VOT/glove
VOT dataset gives the bounding box as float, it can be used directly for ECO and GOTURN, the others needs to be converted to int.
Opencv and Caffe all use BGR color, Matlab use RGB color order.
For timer double timercv = (double)getTickCount();, it should be double, float is not enough.
For frame.copyTo(frameDraw);, only this way, it copy the memory, while frameDraw = frame does not.
For cv::merge(mv, dst), mv should be 1 channel, if more than 1 channel, it has potential problems. Check the code of feature_extractor.cpp\get_cnn_layers\input_channels.