def update_tracker(self, pos, size):
self.base_target_sz = size
# Window size, taking padding into account
self.window_sz = np.floor(
np.array((max(self.base_target_sz), max(self.base_target_sz))) *
(1 + self.parameters.padding))
sz = self.window_sz
sz = np.floor(sz / self.parameters.cell_size)
self.l1_patch_num = np.floor(self.window_sz /
self.parameters.cell_size)
# Desired translation filter output (2d gaussian shaped), bandwidth
# Proportional to target size
output_sigma = np.sqrt(
np.prod(self.base_target_sz)
) * self.parameters.output_sigma_factor / self.parameters.cell_size
self.yf = np.fft.fft2(
desiredResponse.gaussian_response_2d(output_sigma,
self.l1_patch_num))
# Cosine window with the size of the translation filter (2D)
self.cos_window = np.dot(
np.hanning(self.yf.shape[0]).reshape(self.yf.shape[0], 1),
np.hanning(self.yf.shape[1]).reshape(1, self.yf.shape[1]))
self.pos = pos
def __init__(self, im, params):
self.parameters = params
self.pos = self.parameters.init_pos
self.target_sz = self.parameters.target_size
self.graph = None
# Initial target size
self.init_target_sz = self.parameters.target_size
# target sz at scale = 1
self.base_target_sz = self.parameters.target_size
# Window size, taking padding into account
self.window_sz = np.floor(np.array((max(self.base_target_sz),
max(self.base_target_sz))) * (1 + self.parameters.padding))
sz = self.window_sz
sz = np.floor(sz / self.parameters.cell_size)
self.l1_patch_num = np.floor(self.window_sz / self.parameters.cell_size)
# Desired translation filter output (2d gaussian shaped), bandwidth
# Proportional to target size
output_sigma = np.sqrt(np.prod(self.base_target_sz)) * self.parameters.output_sigma_factor / self.parameters.cell_size
self.yf = np.fft.fft2(desiredResponse.gaussian_response_2d(output_sigma, self.l1_patch_num))
# Desired output of scale filter (1d gaussian shaped)
scale_sigma = self.parameters.number_of_scales / np.sqrt(self.parameters.number_of_scales) * self.parameters.scale_sigma_factor
self.ysf = np.fft.fft(desiredResponse.gaussian_response_1d(scale_sigma, self.parameters.number_of_scales))
# Cosine window with the size of the translation filter (2D)
self.cos_window = np.dot(np.hanning(self.yf.shape[0]).reshape(self.yf.shape[0], 1),
np.hanning(self.yf.shape[1]).reshape(1, self.yf.shape[1]))
# Cosine window with the size of the scale filter (1D)
if np.mod(self.parameters.number_of_scales, 2) == 0:
self.scale_window = np.single(np.hanning(self.parameters.number_of_scales + 1))
self.scale_window = self.scale_window[1:]
else:
self.scale_window = np.single(np.hanning(self.parameters.number_of_scales))
# Scale Factors [...0.98 1 1.02 1.0404 ...] NOTE: it is not a incremental value (see the scaleFactors values)
ss = np.arange(1, self.parameters.number_of_scales + 1)
self.scale_factors = self.parameters.scale_step**(np.ceil(self.parameters.number_of_scales / 2.0) - ss)
# If the target size is over the threshold then downsample
if np.prod(self.init_target_sz) > self.parameters.scale_model_max_area:
self.scale_model_factor = np.sqrt(self.parameters.scale_model_max_area/np.prod(self.init_target_sz))
else:
self.scale_model_factor = 1
self.scale_model_sz = np.floor(self.init_target_sz*self.scale_model_factor)
self.currentScaleFactor = 1
self.min_scale_factor = self.parameters.scale_step**np.ceil(np.log(np.max(5.0 / sz)) / np.log(self.parameters.scale_step))
self.max_scale_factor = self.parameters.scale_step**np.floor(np.log(np.min(im.shape[0:-1] / self.base_target_sz)) / np.log(self.parameters.scale_step))
self.confidence = np.array(())
self.high_freq_energy = np.array(())
self.psr = np.array(())
# Flag that indicates if the track lost the target or not
self.lost = False
self.model_alphaf = None
self.model_xf = None
self.sf_den = None
self.sf_num = None
def __init__(self, im, params):
self.parameters = params
self.pos = self.parameters.init_pos
self.target_sz = self.parameters.target_size
if self.parameters.features == 'CNN':
caffe.set_device(0)
caffe.set_mode_gpu()
caffe_root = os.environ['CAFFE_ROOT'] +'/'
import sys
sys.path.insert(0, caffe_root + 'python')
self.net = caffe.Net(caffe_root + 'models/3785162f95cd2d5fee77/VGG_ILSVRC_19_layers_deploy.prototxt',
caffe_root + 'models/3785162f95cd2d5fee77/VGG_ILSVRC_19_layers.caffemodel',
caffe.TEST)
# input preprocessing: 'data' is the name of the input blob == net.inputs[0]
self.transformer = caffe.io.Transformer({'data': self.net.blobs['data'].data.shape})
self.transformer.set_transpose('data', (2, 0, 1))
self.transformer.set_mean('data',
np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))
self.transformer.set_raw_scale('data', 255) # model operates on images in [0,255] range instead of [0,1]
self.transformer.set_channel_swap('data', (0, 1, 2)) # model channels in BGR order instead of RGB
# if self.target_sz[0] / 2.0 > self.target_sz[1]:
# self.target_sz[1] = self.target_sz[0]
# self.pos[1] = self.pos[1] - (self.target_sz[0] - self.target_sz[1]) / 2
# elif self.target_sz[0] < self.target_sz[1]/2.0:
# self.target_sz[0] = self.target_sz[1]
# self.pos[0] = self.pos[0] - (self.target_sz[1] - self.target_sz[0]) / 2
# Initial target size
self.init_target_sz = self.parameters.target_size
# target sz at scale = 1
self.base_target_sz = self.parameters.target_size
# Window size, taking padding into account
self.window_sz = np.floor(np.array((max(self.base_target_sz),
max(self.base_target_sz))) * (1 + self.parameters.padding))
sz = self.window_sz
sz = np.floor(sz / self.parameters.cell_size)
self.l1_patch_num = np.floor(self.window_sz / self.parameters.cell_size)
# Desired translation filter output (2d gaussian shaped), bandwidth
# Proportional to target size
output_sigma = np.sqrt(np.prod(self.base_target_sz)) * self.parameters.output_sigma_factor / self.parameters.cell_size
self.yf = np.fft.fft2(desiredResponse.gaussian_response_2d(output_sigma, self.l1_patch_num))
# Desired output of scale filter (1d gaussian shaped)
scale_sigma = self.parameters.number_of_scales / np.sqrt(self.parameters.number_of_scales) * self.parameters.scale_sigma_factor
self.ysf = np.fft.fft(desiredResponse.gaussian_response_1d(scale_sigma, self.parameters.number_of_scales))
# Cosine window with the size of the translation filter (2D)
self.cos_window = np.dot(np.hanning(self.yf.shape[0]).reshape(self.yf.shape[0], 1),
np.hanning(self.yf.shape[1]).reshape(1, self.yf.shape[1]))
# Cosine window with the size of the scale filter (1D)
if np.mod(self.parameters.number_of_scales, 2) == 0:
self.scale_window = np.single(np.hanning(self.parameters.number_of_scales + 1))
self.scale_window = self.scale_window[1:]
else:
self.scale_window = np.single(np.hanning(self.parameters.number_of_scales))
# Scale Factors [...0.98 1 1.02 1.0404 ...] NOTE: it is not a incremental value (see the scaleFactors values)
ss = np.arange(1, self.parameters.number_of_scales + 1)
self.scale_factors = self.parameters.scale_step**(np.ceil(self.parameters.number_of_scales / 2.0) - ss)
# If the target size is over the threshold then downsample
if np.prod(self.init_target_sz) > self.parameters.scale_model_max_area:
self.scale_model_factor = np.sqrt(self.parameters.scale_model_max_area/np.prod(self.init_target_sz))
else:
self.scale_model_factor = 1
self.scale_model_sz = np.floor(self.init_target_sz*self.scale_model_factor)
self.currentScaleFactor = 1
self.min_scale_factor = self.parameters.scale_step**np.ceil(np.log(np.max(5.0 / sz)) / np.log(self.parameters.scale_step))
self.max_scale_factor = self.parameters.scale_step**np.floor(np.log(np.min(im.shape[0:-1] / self.base_target_sz)) / np.log(self.parameters.scale_step))
self.confidence = np.array(())
self.high_freq_energy = np.array(())
self.psr = np.array(())
# Flag that indicates if the track lost the target or not
self.lost = False
self.model_alphaf = None
self.model_xf = None
self.sf_den = None
self.sf_num = None