The movie "Fury" give me very impression. Tank is one of my favourite vehicles. I intend to estimate the pose of tanks to deepen understanding about the modes in opencv. So I tried some tracking algorithm to find the target in the video. And use some cascades to detect different poseture in the tracking box. Maybe it is not a good scheme.
In some special cases the taret is big in the video and I select many images in these cases to test whether the cascade could detect the pose of tank by itself. The target is about 400*400 in 1080*720 images. I adjust the size of all images to 100*80. And use the tools to make 30 positive samples (with the same posture) and 1366 negative samples. In the 4 stage the required leaf false alarm rate achieved. However, the XML file can not work well. It could not detect the traget in images which contain positive samples. The LBP and HOG can not describe the tank? Or something wrong in my training process.
How about trying the "Real Time pose estimation of a textured object" in opencv 3.0.0dec. Could anyone give me some directions?