def track(self, image):
self.frame_num += 1
# Convert image
im = numpy_to_torch(image)
# ------- LOCALIZATION ------- #
# Get sample
sample_pos = self.pos.round()
sample_scales = self.target_scale * self.params.scale_factors
test_xf = self.extract_fourier_sample(im, self.pos, sample_scales,
self.img_sample_sz)
# Compute scores
sf = self.apply_filter(test_xf)
translation_vec, scale_ind, s = self.localize_target(sf)
scale_change_factor = self.params.scale_factors[scale_ind]
# Update position and scale
self.update_state(sample_pos + translation_vec,
self.target_scale * scale_change_factor)
self.predict_target_box(sample_pos, sample_scales[scale_ind],
scale_ind)
if self.params.debug >= 2:
show_tensor(s[scale_ind, ...], 5)
if self.params.debug >= 3:
for i, hf in enumerate(self.filter):
show_tensor(fourier.sample_fs(hf).abs().mean(1), 6 + i)
# ------- UPDATE ------- #
# Get train sample
train_xf = TensorList(
[xf[scale_ind:scale_ind + 1, ...] for xf in test_xf])
# Shift the sample
shift_samp = 2 * math.pi * (self.pos - sample_pos) / (
sample_scales[scale_ind] * self.img_support_sz)
train_xf = fourier.shift_fs(train_xf, shift=shift_samp)
# Update memory
self.update_memory(train_xf)
# Train filter
if self.frame_num % self.params.train_skipping == 1:
self.filter_optimizer.run(self.params.CG_iter, train_xf)
self.symmetrize_filter()
# Return new state
new_state = torch.cat(
(self.pos[[1, 0]] - (self.target_sz[[1, 0]] - 1) / 2,
self.target_sz[[1, 0]]))
return new_state.tolist()
def localize_target(self, scores_raw):
if self.params.score_fusion_strategy == 'weightedsum':
weight = self.fparams.attribute('translation_weight')
scores_raw = weight * scores_raw
sf_weighted = fourier.cfft2(scores_raw) / (scores_raw.size(2) *
scores_raw.size(3))
for i, (sz,
ksz) in enumerate(zip(self.feature_sz, self.kernel_size)):
sf_weighted[i] = fourier.shift_fs(
sf_weighted[i],
math.pi *
(1 - torch.Tensor([ksz[0] % 2, ksz[1] % 2]) / sz))
scores_fs = fourier.sum_fs(sf_weighted)
scores = fourier.sample_fs(scores_fs, self.output_sz)
elif self.params.score_fusion_strategy == 'default':
if len(scores_raw) > 1:
raise NotImplementedError('Not implemented')
scores = scores_raw[0]
ksz = self.kernel_size[0]
offset = torch.Tensor([ksz[0] % 2, ksz[1] % 2]) / 2
else:
raise ValueError('Unknown score fusion strategy.')
if self.output_window is not None and not getattr(
self.params, 'perform_hn_without_windowing', False):
raise NotImplementedError
scores *= self.output_window
if getattr(self.params, 'advanced_localization', False):
return self.localize_advanced(scores)
# Get maximum
max_score, max_disp = dcf.max2d(scores)
_, scale_ind = torch.max(max_score, dim=0)
max_disp = max_disp.float().cpu()
# Convert to displacements in the base scale
if self.params.score_fusion_strategy == 'default':
disp = max_disp + offset
else:
disp = (max_disp +
self.output_sz / 2) % self.output_sz - self.output_sz / 2
# Compute translation vector and scale change factor
translation_vec = disp[scale_ind, ...].view(-1) * (
self.img_support_sz / self.output_sz) * self.target_scale
translation_vec *= self.params.scale_factors[scale_ind]
# Shift the score output for visualization purposes
if self.params.debug >= 2:
sz = scores.shape[-2:]
scores = torch.cat(
[scores[..., sz[0] // 2:, :], scores[..., :sz[0] // 2, :]], -2)
scores = torch.cat(
[scores[..., :, sz[1] // 2:], scores[..., :, :sz[1] // 2]], -1)
return translation_vec, scale_ind, scores, None
def localize_target(self, scores_raw):
# Weighted sum (if multiple features) with interpolation in fourier domain
weight = self.fparams.attribute('translation_weight',
1.0) #weight 没什么用
if (Debug):
print("weight : ", weight) #
scores_raw = weight * scores_raw #
if (Debug):
print("scores_raw: ", scores_raw)
sf_weighted = fourier.cfft2(scores_raw) / (scores_raw.size(2) *
scores_raw.size(3))
for i, (sz, ksz) in enumerate(zip(self.feature_sz, self.kernel_size)):
# """Shift a sample a in the Fourier domain.
sf_weighted[i] = fourier.shift_fs(
sf_weighted[i],
math.pi * (1 - torch.Tensor([ksz[0] % 2, ksz[1] % 2]) / sz))
#"""Sum a list of Fourier series expansions."""
scores_fs = fourier.sum_fs(sf_weighted)
if (Debug):
print("scores_fs : ", scores_fs)
#"""Samples the Fourier series."""
scores = fourier.sample_fs(scores_fs, self.output_sz)
if (Debug):
print("scores: ", scores)
if self.output_window is not None and not getattr(
self.params, 'perform_hn_without_windowing', False):
scores *= self.output_window
if getattr(self.params, 'advanced_localization', False):
if (Debug):
print("advanced: ")
return self.localize_advanced(scores)
# Get maximum
max_score, max_disp = dcf.max2d(scores)
_, scale_ind = torch.max(max_score, dim=0)
max_disp = max_disp.float().cpu()
# Convert to displacements in the base scale
disp = (max_disp +
self.output_sz / 2) % self.output_sz - self.output_sz / 2
# Compute translation vector and scale change factor
translation_vec = disp[scale_ind, ...].view(-1) * (
self.img_support_sz / self.output_sz) * self.target_scale
translation_vec *= self.params.scale_factors[scale_ind]
# Shift the score output for visualization purposes
if self.params.debug >= 2:
sz = scores.shape[-2:]
scores = torch.cat(
[scores[..., sz[0] // 2:, :], scores[..., :sz[0] // 2, :]], -2)
scores = torch.cat(
[scores[..., :, sz[1] // 2:], scores[..., :, :sz[1] // 2]], -1)
return translation_vec, scale_ind, scores, None
def localize_target(self, sf: TensorList):
if self.params.score_fusion_strategy == 'sum':
scores = fourier.sample_fs(fourier.sum_fs(sf), self.output_sz)
elif self.params.score_fusion_strategy == 'weightedsum':
weight = self.fparams.attribute('translation_weight')
scores = fourier.sample_fs(fourier.sum_fs(weight * sf),
self.output_sz)
elif self.params.score_fusion_strategy == 'transcale':
alpha = self.fparams.attribute('scale_weight')
beta = self.fparams.attribute('translation_weight')
sample_sz = torch.round(
self.output_sz.view(1, -1) *
self.params.scale_factors.view(-1, 1))
scores = 0
for sfe, a, b in zip(sf, alpha, beta):
sfe = fourier.shift_fs(sfe, math.pi * torch.ones(2))
scores_scales = []
for sind, sz in enumerate(sample_sz):
pd = (self.output_sz - sz) / 2
scores_scales.append(
F.pad(fourier.sample_fs(sfe[sind:sind + 1, ...], sz),
(math.floor(pd[1].item()), math.ceil(
pd[1].item()), math.floor(
pd[0].item()), math.ceil(pd[0].item()))))
scores_cat = torch.cat(scores_scales)
scores = scores + (b - a) * scores_cat.mean(
dim=0, keepdim=True) + a * scores_cat
else:
raise ValueError('Unknown score fusion strategy.')
# Get maximum
max_score, max_disp = dcf.max2d(scores)
_, scale_ind = torch.max(max_score, dim=0)
max_disp = max_disp.float().cpu()
# Convert to displacements in the base scale
if self.params.score_fusion_strategy in ['sum', 'weightedsum']:
disp = (max_disp +
self.output_sz / 2) % self.output_sz - self.output_sz / 2
elif self.params.score_fusion_strategy == 'transcale':
disp = max_disp - self.output_sz / 2
# Compute translation vector and scale change factor
translation_vec = disp[scale_ind, ...].view(-1) * (
self.img_support_sz / self.output_sz) * self.target_scale
if self.params.score_fusion_strategy in ['sum', 'weightedsum']:
translation_vec *= self.params.scale_factors[scale_ind]
return translation_vec, scale_ind, scores
def initialize(self, image, info: dict) -> dict:
initSeed = 1
torch.manual_seed(initSeed)
torch.cuda.manual_seed(initSeed)
torch.cuda.manual_seed_all(initSeed)
np.random.seed(initSeed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
os.environ['PYTHONHASHSEED'] = str(initSeed)
state = info['init_bbox']
# Initialize some stuff
self.frame_num = 1
if not hasattr(self.params, 'device'):
self.params.device = 'cuda' if self.params.use_gpu else 'cpu'
# Initialize features
self.initialize_features()
# metricnet
self.metric_model = model_load(self.params.metric_model_path)
# warmup start
with torch.no_grad():
tmp = np.random.rand(5, 3, 107, 107)
tmp = torch.Tensor(tmp)
tmp = (Variable(tmp)).type(torch.FloatTensor).cuda()
tmp = self.metric_model(tmp)
# warmup end
self.target_metric_feature = get_target_feature(
self.metric_model, np.array(state), np.array(image))
pos_generator = SampleGenerator(
'gaussian', np.array([image.shape[1], image.shape[0]]), 0.1, 1.3)
gt_pos_examples = pos_generator(
np.array(state).astype(np.int), 20, [0.7, 1])
gt_iou = 0.7
while gt_pos_examples.shape[0] == 0:
gt_iou = gt_iou - 0.1
gt_pos_examples = pos_generator(
np.array(state).astype(np.int), 20, [gt_iou, 1])
# print('gt-iou：', gt_iou)
# self.gt_pos_features = get_anchor_feature(self.metric_model, np.array(image), gt_pos_examples).cpu().detach().numpy()
with torch.no_grad():
gt_pos_features0 = get_anchor_feature(self.metric_model,
np.array(image),
gt_pos_examples)
gt_pos_features = gt_pos_features0.cpu().detach().numpy()
target_metric_feature = self.target_metric_feature.repeat(
gt_pos_features.shape[0], 1)
pos_all = torch.norm(gt_pos_features0 - target_metric_feature,
2,
dim=1).view(-1)
self.similar = pos_all.mean() * self.params.sim_rate
print('similarThresh', self.similar)
self.target_features_all = []
self.target_features_all.append(self.target_metric_feature)
self.clf = lof_fit(gt_pos_features, k=5)
# Chack if image is color
self.params.features.set_is_color(image.shape[2] == 3)
# Get feature specific params
self.fparams = self.params.features.get_fparams('feature_params')
# Get position and size
self.pos = torch.Tensor(
[state[1] + (state[3] - 1) / 2, state[0] + (state[2] - 1) / 2])
self.target_sz = torch.Tensor([state[3], state[2]])
# Set search area
self.target_scale = 1.0
search_area = torch.prod(self.target_sz *
self.params.search_area_scale).item()
if search_area > self.params.max_image_sample_size:
self.target_scale = math.sqrt(search_area /
self.params.max_image_sample_size)
elif search_area < self.params.min_image_sample_size:
self.target_scale = math.sqrt(search_area /
self.params.min_image_sample_size)
# Target size in base scale
self.base_target_sz = self.target_sz / self.target_scale
# Use odd square search area and set sizes
feat_max_stride = max(self.params.features.stride())
self.img_sample_sz = torch.round(
torch.sqrt(
torch.prod(self.base_target_sz *
self.params.search_area_scale))) * torch.ones(2)
self.img_sample_sz += feat_max_stride - self.img_sample_sz % (
2 * feat_max_stride)
# Set other sizes (corresponds to ECO code)
self.img_support_sz = self.img_sample_sz
self.feature_sz = self.params.features.size(self.img_sample_sz)
self.filter_sz = self.feature_sz + (self.feature_sz + 1) % 2
self.output_sz = self.params.score_upsample_factor * self.img_support_sz # Interpolated size of the output
self.compressed_dim = self.fparams.attribute('compressed_dim')
# Number of filters
self.num_filters = len(self.filter_sz)
# Get window function
self.window = TensorList(
[dcf.hann2d(sz).to(self.params.device) for sz in self.feature_sz])
# Get interpolation function
self.interp_fs = TensorList([
dcf.get_interp_fourier(sz, self.params.interpolation_method,
self.params.interpolation_bicubic_a,
self.params.interpolation_centering,
self.params.interpolation_windowing,
self.params.device) for sz in self.filter_sz
])
# Get regularization filter
self.reg_filter = TensorList([
dcf.get_reg_filter(self.img_support_sz, self.base_target_sz,
fparams).to(self.params.device)
for fparams in self.fparams
])
self.reg_energy = self.reg_filter.view(-1) @ self.reg_filter.view(-1)
# Get label function
output_sigma_factor = self.fparams.attribute('output_sigma_factor')
sigma = (self.filter_sz / self.img_support_sz) * torch.sqrt(
self.base_target_sz.prod()) * output_sigma_factor
self.yf = TensorList([
dcf.label_function(sz, sig).to(self.params.device)
for sz, sig in zip(self.filter_sz, sigma)
])
# Optimization options
self.params.precond_learning_rate = self.fparams.attribute(
'learning_rate')
if self.params.CG_forgetting_rate is None or max(
self.params.precond_learning_rate) >= 1:
self.params.direction_forget_factor = 0
else:
self.params.direction_forget_factor = (
1 - max(self.params.precond_learning_rate)
)**self.params.CG_forgetting_rate
# Convert image
im = numpy_to_torch(image)
# Setup bounds
self.image_sz = torch.Tensor([im.shape[2], im.shape[3]])
self.min_scale_factor = torch.max(10 / self.base_target_sz)
self.max_scale_factor = torch.min(self.image_sz / self.base_target_sz)
# Extract and transform sample
x = self.generate_init_samples(im)
# Initialize projection matrix
x_mat = TensorList(
[e.permute(1, 0, 2, 3).reshape(e.shape[1], -1).clone() for e in x])
x_mat -= x_mat.mean(dim=1, keepdim=True)
cov_x = x_mat @ x_mat.t()
self.projection_matrix = TensorList([
torch.svd(C)[0][:, :cdim].clone()
for C, cdim in zip(cov_x, self.compressed_dim)
])
# Transform to get the training sample
train_xf = self.preprocess_sample(x)
# Shift the samples back
if 'shift' in self.params.augmentation:
for xf in train_xf:
if xf.shape[0] == 1:
continue
for i, shift in enumerate(self.params.augmentation['shift']):
shift_samp = 2 * math.pi * torch.Tensor(
shift) / self.img_support_sz
xf[1 + i:2 + i, ...] = fourier.shift_fs(xf[1 + i:2 + i,
...],
shift=shift_samp)
# Shift sample
shift_samp = 2 * math.pi * (self.pos - self.pos.round()) / (
self.target_scale * self.img_support_sz)
train_xf = fourier.shift_fs(train_xf, shift=shift_samp)
# Initialize first-frame training samples
num_init_samples = train_xf.size(0)
self.init_sample_weights = TensorList(
[xf.new_ones(1) / xf.shape[0] for xf in train_xf])
self.init_training_samples = train_xf.permute(2, 3, 0, 1, 4)
# Sample counters and weights
self.num_stored_samples = num_init_samples
self.previous_replace_ind = [None] * len(self.num_stored_samples)
self.sample_weights = TensorList(
[xf.new_zeros(self.params.sample_memory_size) for xf in train_xf])
for sw, init_sw, num in zip(self.sample_weights,
self.init_sample_weights,
num_init_samples):
sw[:num] = init_sw
# Initialize memory
self.training_samples = TensorList([
xf.new_zeros(xf.shape[2], xf.shape[3],
self.params.sample_memory_size, cdim, 2)
for xf, cdim in zip(train_xf, self.compressed_dim)
])
# Initialize filter
self.filter = TensorList([
xf.new_zeros(1, cdim, xf.shape[2], xf.shape[3], 2)
for xf, cdim in zip(train_xf, self.compressed_dim)
])
# Do joint optimization
self.joint_problem = FactorizedConvProblem(self.init_training_samples,
self.yf, self.reg_filter,
self.projection_matrix,
self.params,
self.init_sample_weights)
joint_var = self.filter.concat(self.projection_matrix)
self.joint_optimizer = GaussNewtonCG(self.joint_problem,
joint_var,
debug=(self.params.debug >= 1),
visdom=self.visdom)
if self.params.update_projection_matrix:
self.joint_optimizer.run(
self.params.init_CG_iter // self.params.init_GN_iter,
self.params.init_GN_iter)
# Re-project samples with the new projection matrix
compressed_samples = complex.mtimes(self.init_training_samples,
self.projection_matrix)
for train_samp, init_samp in zip(self.training_samples,
compressed_samples):
train_samp[:, :, :init_samp.shape[2], :, :] = init_samp
# Initialize optimizer
self.filter_optimizer = FilterOptim(self.params, self.reg_energy)
self.filter_optimizer.register(self.filter, self.training_samples,
self.yf, self.sample_weights,
self.reg_filter)
self.filter_optimizer.sample_energy = self.joint_problem.sample_energy
self.filter_optimizer.residuals = self.joint_optimizer.residuals.clone(
)
if not self.params.update_projection_matrix:
self.filter_optimizer.run(self.params.init_CG_iter)
# Post optimization
self.filter_optimizer.run(self.params.post_init_CG_iter)
self.symmetrize_filter()
# metricnet_lof
self.current_target_metric_feature = []
self.train_xf = []
# self.iou=[]
# self.lof_thresh=3.5
self.lof_thresh = self.params.lof_rate
def track(self, image) -> dict:
self.debug_info = {}
self.frame_num += 1
self.debug_info['frame_num'] = self.frame_num
# Convert image
im = numpy_to_torch(image)
# ------- LOCALIZATION ------- #
# Get sample
sample_pos = self.pos.round()
sample_scales = self.target_scale * self.params.scale_factors
test_xf = self.extract_fourier_sample(im, self.pos, sample_scales,
self.img_sample_sz)
# Compute scores
sf = self.apply_filter(test_xf)
translation_vec, scale_ind, s = self.localize_target(sf)
scale_change_factor = self.params.scale_factors[scale_ind]
# Update position and scale
self.update_state(sample_pos + translation_vec,
self.target_scale * scale_change_factor)
score_map = s[scale_ind, ...]
max_score = torch.max(score_map).item()
self.debug_info['max_score'] = max_score
if self.visdom is not None:
self.visdom.register(score_map, 'heatmap', 2, 'Score Map')
self.visdom.register(self.debug_info, 'info_dict', 1, 'Status')
elif self.params.debug >= 2:
show_tensor(score_map,
5,
title='Max score = {:.2f}'.format(max_score))
# if self.params.debug >= 3:
# for i, hf in enumerate(self.filter):
# show_tensor(fourier.sample_fs(hf).abs().mean(1), 6+i)
# metric
state_tmp = torch.cat(
(self.pos[[1, 0]] - (self.target_sz[[1, 0]] - 1) / 2,
self.target_sz[[1, 0]]))
state_tmp = state_tmp.numpy()
with torch.no_grad():
self.current_target_metric_feature.append(
get_target_feature(self.metric_model, state_tmp,
np.array(image)).cpu().detach().numpy())
# self.iou.append(overlap_ratio(state_tmp,self.ground_truth_rect[self.frame_num-1]))
# success, target_dist = judge_success_no_class(self.metric_model, current_target_metric_feature,self.target_metric_feature, self.params)
# lof_predict,success = lof(self.gt_pos_features, current_target_metric_feature.cpu().detach().numpy().reshape((1,1024)), k=5,thresh=5)
# print(self.frame_num,': lof:',lof_predict[0],' ',success[0])
# ------- UPDATE ------- #
# Get train sample
train_xf = TensorList(
[xf[scale_ind:scale_ind + 1, ...] for xf in test_xf])
# Shift the sample
shift_samp = 2 * math.pi * (self.pos - sample_pos) / (
sample_scales[scale_ind] * self.img_support_sz)
train_xf = fourier.shift_fs(train_xf, shift=shift_samp)
self.train_xf.append(train_xf)
if self.frame_num == 1:
# Update memory
self.update_memory(train_xf) # metricnet
self.filter_optimizer.run(self.params.CG_iter, train_xf)
self.symmetrize_filter()
elif self.frame_num % self.params.train_skipping == 1:
current_target_metric_feature = np.array(
self.current_target_metric_feature).squeeze()
current_target_metric_feature0 = torch.from_numpy(
current_target_metric_feature).cuda()
# lof_predict, success = lof(np.concatenate([self.gt_pos_features,current_target_metric_feature],axis=0), k=20,thresh=self.lof_thresh)
lof_predict, success = lof(current_target_metric_feature,
self.clf,
k=5,
thresh=self.lof_thresh)
last_id = -1
if self.frame_num <= self.params.train_skipping + 1:
self.lof_thresh = lof_predict.mean() * self.params.lof_rate
print('lof_thresh:', self.lof_thresh)
for ii in range(len(self.train_xf)):
# print('lof:',lof_predict[ii],' iou:',self.iou[ii],success[ii])
if self.frame_num > self.params.train_skipping + 1 and success[
ii]:
for kk in range(len(self.target_features_all) - 1, -1, -1):
dist = torch.norm(
self.target_features_all[kk] -
current_target_metric_feature0[ii].reshape(
[1, 1024]),
2,
dim=1).view(-1)
if dist < self.similar:
success[ii] = 0
continue
if self.frame_num <= self.params.train_skipping + 1 or success[
ii]:
self.target_features_all.append(
current_target_metric_feature0[ii].reshape([1, 1024]))
last_id = ii
self.update_memory(self.train_xf[ii])
if last_id > -1:
self.filter_optimizer.run(self.params.CG_iter,
self.train_xf[last_id])
self.symmetrize_filter()
self.current_target_metric_feature = []
self.train_xf = []
# self.iou=[]
# # Train filter
# if self.frame_num % self.params.train_skipping == 1:
# self.filter_optimizer.run(self.params.CG_iter, train_xf)
# self.symmetrize_filter()
# Return new state
new_state = torch.cat(
(self.pos[[1, 0]] - (self.target_sz[[1, 0]] - 1) / 2,
self.target_sz[[1, 0]]))
out = {'target_bbox': new_state.tolist()}
return out
def initialize(self, image, state, *args, **kwargs):
# Initialize some stuff
self.frame_num = 1
if not hasattr(self.params, 'device'):
self.params.device = 'cuda' if self.params.use_gpu else 'cpu'
# Initialize features
self.initialize_features()
# Chack if image is color
self.params.features.set_is_color(image.shape[2] == 3)
# Get feature specific params
self.fparams = self.params.features.get_fparams('feature_params')
# Get position and size
self.pos = torch.Tensor([state[1] + (state[3] - 1)/2, state[0] + (state[2] - 1)/2])
self.target_sz = torch.Tensor([state[3], state[2]])
# Set search area
self.target_scale = 1.0
search_area = torch.prod(self.target_sz * self.params.search_area_scale).item()
if search_area > self.params.max_image_sample_size:
self.target_scale = math.sqrt(search_area / self.params.max_image_sample_size)
elif search_area < self.params.min_image_sample_size:
self.target_scale = math.sqrt(search_area / self.params.min_image_sample_size)
# Target size in base scale
self.base_target_sz = self.target_sz / self.target_scale
# Use odd square search area and set sizes
feat_max_stride = max(self.params.features.stride())
self.img_sample_sz = torch.round(torch.sqrt(torch.prod(self.base_target_sz * self.params.search_area_scale))) * torch.ones(2)
self.img_sample_sz += feat_max_stride - self.img_sample_sz % (2 * feat_max_stride)
# Set other sizes (corresponds to ECO code)
self.img_support_sz = self.img_sample_sz
self.feature_sz = self.params.features.size(self.img_sample_sz)
self.filter_sz = self.feature_sz + (self.feature_sz + 1) % 2
self.output_sz = self.params.score_upsample_factor * self.img_support_sz # Interpolated size of the output
self.compressed_dim = self.fparams.attribute('compressed_dim')
# Number of filters
self.num_filters = len(self.filter_sz)
# Get window function
self.window = TensorList([dcf.hann2d(sz).to(self.params.device) for sz in self.feature_sz])
# Get interpolation function
self.interp_fs = TensorList([dcf.get_interp_fourier(sz, self.params.interpolation_method,
self.params.interpolation_bicubic_a, self.params.interpolation_centering,
self.params.interpolation_windowing, self.params.device) for sz in self.filter_sz])
# Get regularization filter
self.reg_filter = TensorList([dcf.get_reg_filter(self.img_support_sz, self.base_target_sz, fparams).to(self.params.device)
for fparams in self.fparams])
self.reg_energy = self.reg_filter.view(-1) @ self.reg_filter.view(-1)
# Get label function
output_sigma_factor = self.fparams.attribute('output_sigma_factor')
sigma = (self.filter_sz / self.img_support_sz) * torch.sqrt(self.base_target_sz.prod()) * output_sigma_factor
self.yf = TensorList([dcf.label_function(sz, sig).to(self.params.device) for sz, sig in zip(self.filter_sz, sigma)])
# Optimization options
self.params.precond_learning_rate = self.fparams.attribute('learning_rate')
if self.params.CG_forgetting_rate is None or max(self.params.precond_learning_rate) >= 1:
self.params.direction_forget_factor = 0
else:
self.params.direction_forget_factor = (1 - max(self.params.precond_learning_rate))**self.params.CG_forgetting_rate
# Convert image
im = numpy_to_torch(image)
# Setup bounds
self.image_sz = torch.Tensor([im.shape[2], im.shape[3]])
self.min_scale_factor = torch.max(10 / self.base_target_sz)
self.max_scale_factor = torch.min(self.image_sz / self.base_target_sz)
# Extract and transform sample
x = self.generate_init_samples(im)
# Initialize projection matrix
x_mat = TensorList([e.permute(1,0,2,3).reshape(e.shape[1], -1).clone() for e in x])
x_mat -= x_mat.mean(dim=1, keepdim=True)
cov_x = x_mat @ x_mat.t()
self.projection_matrix = TensorList([torch.svd(C)[0][:,:cdim].clone() for C, cdim in zip(cov_x, self.compressed_dim)])
# Transform to get the training sample
train_xf = self.preprocess_sample(x)
# Shift the samples back
if 'shift' in self.params.augmentation:
for xf in train_xf:
if xf.shape[0] == 1:
continue
for i, shift in enumerate(self.params.augmentation['shift']):
shift_samp = 2 * math.pi * torch.Tensor(shift) / self.img_support_sz
xf[1+i:2+i,...] = fourier.shift_fs(xf[1+i:2+i,...], shift=shift_samp)
# Shift sample
shift_samp = 2*math.pi * (self.pos - self.pos.round()) / (self.target_scale * self.img_support_sz)
train_xf = fourier.shift_fs(train_xf, shift=shift_samp)
# Initialize first-frame training samples
num_init_samples = train_xf.size(0)
self.init_sample_weights = TensorList([xf.new_ones(1) / xf.shape[0] for xf in train_xf])
self.init_training_samples = train_xf.permute(2, 3, 0, 1, 4)
# Sample counters and weights
self.num_stored_samples = num_init_samples
self.previous_replace_ind = [None]*len(self.num_stored_samples)
self.sample_weights = TensorList([xf.new_zeros(self.params.sample_memory_size) for xf in train_xf])
for sw, init_sw, num in zip(self.sample_weights, self.init_sample_weights, num_init_samples):
sw[:num] = init_sw
# Initialize memory
self.training_samples = TensorList(
[xf.new_zeros(xf.shape[2], xf.shape[3], self.params.sample_memory_size, cdim, 2) for xf, cdim in zip(train_xf, self.compressed_dim)])
# Initialize filter
self.filter = TensorList(
[xf.new_zeros(1, cdim, xf.shape[2], xf.shape[3], 2) for xf, cdim in zip(train_xf, self.compressed_dim)])
# Do joint optimization
self.joint_problem = FactorizedConvProblem(self.init_training_samples, self.yf, self.reg_filter, self.projection_matrix, self.params, self.init_sample_weights)
joint_var = self.filter.concat(self.projection_matrix)
self.joint_optimizer = GaussNewtonCG(self.joint_problem, joint_var, debug=(self.params.debug>=3))
if self.params.update_projection_matrix:
self.joint_optimizer.run(self.params.init_CG_iter // self.params.init_GN_iter, self.params.init_GN_iter)
# Re-project samples with the new projection matrix
compressed_samples = complex.mtimes(self.init_training_samples, self.projection_matrix)
for train_samp, init_samp in zip(self.training_samples, compressed_samples):
train_samp[:,:,:init_samp.shape[2],:,:] = init_samp
# Initialize optimizer
self.filter_optimizer = FilterOptim(self.params, self.reg_energy)
self.filter_optimizer.register(self.filter, self.training_samples, self.yf, self.sample_weights, self.reg_filter)
self.filter_optimizer.sample_energy = self.joint_problem.sample_energy
self.filter_optimizer.residuals = self.joint_optimizer.residuals.clone()
if not self.params.update_projection_matrix:
self.filter_optimizer.run(self.params.init_CG_iter)
# Post optimization
self.filter_optimizer.run(self.params.post_init_CG_iter)
self.symmetrize_filter()
def initialize(self, image, info: dict, gpu_device) -> dict:
# Initialize some stuff
self.frame_num = 1
self.params.device = 'cuda:{0}'.format(
gpu_device) if self.params.use_gpu else 'cpu'
# Convert image
im = numpy_to_torch(image)
self.image_sz = torch.Tensor([im.shape[2], im.shape[3]])
# Initialize features
self.initialize_features(im)
# Chack if image is color
self.params.features.set_is_color(image.shape[2] == 3)
# Get feature specific params
self.fparams = self.params.features.get_fparams('feature_params')
# Get position and size
self.points = TensorList(
[torch.Tensor([p[0], p[1]]) for p in info['points']])
self.org_points = self.points.clone()
self.target_sz = torch.Tensor(
[info['target_sz'][0], info['target_sz'][1]])
# Use odd square search area and set sizes
feat_max_stride = max(self.params.features.stride())
self.img_sample_sz = self.image_sz.clone()
self.img_sample_sz += feat_max_stride - self.img_sample_sz % (
2 * feat_max_stride)
# Set other sizes (corresponds to ECO code)
self.img_support_sz = self.img_sample_sz
self.mid_point = self.img_support_sz // 2
self.feature_sz = self.params.features.size(self.img_sample_sz)
self.filter_sz = self.feature_sz + (self.feature_sz + 1) % 2
self.output_sz = self.img_support_sz # Interpolated size of the output
# Number of filters
self.num_filters = len(self.filter_sz)
# Get window function
#self.window = TensorList([dcf.hann2d(sz).to(self.params.device) for sz in self.feature_sz])
self.window = TensorList([
torch.ones((1, 1, int(sz[0].item()),
int(sz[1].item()))).to(self.params.device)
for sz in self.feature_sz
])
#self.window = TensorList([dcf.tukey2d(sz).to(self.params.device) for sz in self.feature_sz])
# Get interpolation function
self.interp_fs = TensorList([
dcf.get_interp_fourier(sz, self.params.interpolation_method,
self.params.interpolation_bicubic_a,
self.params.interpolation_centering,
self.params.interpolation_windowing,
self.params.device) for sz in self.filter_sz
])
# Get label function
output_sigma_factor = self.fparams.attribute('output_sigma_factor')
sigma = (self.filter_sz / self.img_support_sz) * torch.sqrt(
self.target_sz.prod()) * output_sigma_factor
yf_zero = TensorList([
dcf.label_function(sz, sig).to(self.params.device)
for sz, sig in zip(self.filter_sz, sigma)
])
yf_zero = complex.complex(yf_zero)
self.yf = TensorList()
for p in self.points:
shift_sample = 2 * math.pi * (self.mid_point -
p) / self.img_support_sz
self.yf.append(
TensorList(
[fourier.shift_fs(yfs, shift_sample) for yfs in yf_zero]))
# Optimization options
self.params.precond_learning_rate = self.fparams.attribute(
'learning_rate')
if self.params.CG_forgetting_rate is None or max(
self.params.precond_learning_rate) >= 1:
self.params.direction_forget_factor = 0
else:
self.params.direction_forget_factor = (
1 - max(self.params.precond_learning_rate)
)**self.params.CG_forgetting_rate
# Extract and transform sample
x = self.generate_init_samples(im).to(self.params.device)
self.x = x
# Transform to get the training sample
train_xf = self.preprocess_sample(x)
# Shift the samples back
if 'shift' in self.params.augmentation:
for xf in train_xf:
if xf.shape[0] == 1:
continue
for i, shift in enumerate(self.params.augmentation['shift']):
shift_samp = 2 * math.pi * torch.Tensor(
shift) / self.img_support_sz
xf[1 + i:2 + i, ...] = fourier.shift_fs(xf[1 + i:2 + i,
...],
shift=shift_samp)
# Initialize first-frame training samples
num_init_samples = train_xf.size(0)
self.init_training_samples = train_xf.permute(2, 3, 0, 1, 4)
# Initialize memory
# Initialize filter
self.training_samples = TensorList([
xf.new_zeros(xf.shape[2], xf.shape[3],
self.params.sample_memory_size, xf.shape[1], 2)
for xf in train_xf
])
self.filters = TensorList([
TensorList([
xf.new_zeros(1, xf.shape[1], xf.shape[2], xf.shape[3], 2)
for xf in train_xf
]) for i in range(len(self.points))
])
self.init_sample_weights = TensorList(
[xf.new_ones(1) / xf.shape[0] for xf in train_xf])
self.sample_weights = TensorList(
[xf.new_zeros(self.params.sample_memory_size) for xf in train_xf])
for sw, init_sw, num in zip(self.sample_weights,
self.init_sample_weights,
num_init_samples):
sw[:num] = init_sw
# Get regularization filter
self.reg_filter = TensorList([
dcf.get_reg_filter(self.img_support_sz, self.target_sz,
fparams).to(self.params.device)
for fparams in self.fparams
])
self.reg_energy = self.reg_filter.view(-1) @ self.reg_filter.view(-1)
# Sample counters and weights
self.num_stored_samples = num_init_samples
self.previous_replace_ind = [None] * len(self.num_stored_samples)
for train_samp, init_samp in zip(self.training_samples,
self.init_training_samples):
train_samp[:, :, :init_samp.shape[2], :, :] = init_samp
sample_energy = complex.abs_sqr(self.training_samples).mean(
dim=2, keepdim=True).permute(2, 3, 0, 1)
# Do joint optimization
for i in range(len(self.points)):
print('{0}'.format(i), end=', ')
ts = self.training_samples.clone()
yf = self.yf[i]
filters = self.filters[i]
i_sw = self.init_sample_weights.clone()
re = self.reg_energy.clone()
sw = self.sample_weights.clone()
rf = self.reg_filter.clone()
filter_optimizer = FilterOptim(self.params, re)
filter_optimizer.register(filters, ts, yf, sw, rf)
filter_optimizer.sample_energy = sample_energy.clone()
filter_optimizer.run(self.params.init_CG_iter)
# Post optimization
filter_optimizer.run(self.params.post_init_CG_iter)
self.filters[i] = filter_optimizer.filter
self.symmetrize_filter()
print()
def track(self, image, info: dict = None) -> dict:
self.debug_info = {}
self.frame_num += 1
self.debug_info['frame_num'] = self.frame_num
# Convert image
im = numpy_to_torch(image)
# ------- LOCALIZATION ------- #
# Get sample
sample_pos = self.pos.round()
sample_scales = self.target_scale * self.params.scale_factors
test_xf = self.extract_fourier_sample(im, self.pos, sample_scales, self.img_sample_sz)
# Compute scores
sf = self.apply_filter(test_xf)
translation_vec, scale_ind, s = self.localize_target(sf)
scale_change_factor = self.params.scale_factors[scale_ind]
# Update position and scale
self.update_state(sample_pos + translation_vec, self.target_scale * scale_change_factor)
score_map = s[scale_ind, ...]
max_score = torch.max(score_map).item()
self.debug_info['max_score'] = max_score
if self.visdom is not None:
self.visdom.register(score_map, 'heatmap', 2, 'Score Map')
self.visdom.register(self.debug_info, 'info_dict', 1, 'Status')
elif self.params.debug >= 2:
show_tensor(score_map, 5, title='Max score = {:.2f}'.format(max_score))
# if self.params.debug >= 3:
# for i, hf in enumerate(self.filter):
# show_tensor(fourier.sample_fs(hf).abs().mean(1), 6+i)
# ------- UPDATE ------- #
# Get train sample
train_xf = TensorList([xf[scale_ind:scale_ind+1, ...] for xf in test_xf])
# Shift the sample
shift_samp = 2*math.pi * (self.pos - sample_pos) / (sample_scales[scale_ind] * self.img_support_sz)
train_xf = fourier.shift_fs(train_xf, shift=shift_samp)
# Update memory
self.update_memory(train_xf)
# Train filter
if self.frame_num % self.params.train_skipping == 1:
self.filter_optimizer.run(self.params.CG_iter, train_xf)
self.symmetrize_filter()
# Return new state
new_state = torch.cat((self.pos[[1,0]] - (self.target_sz[[1,0]]-1)/2, self.target_sz[[1,0]]))
out = {'target_bbox': new_state.tolist()}
return out
