def visual_batch_eval(cam, image, length, k):
cam = torch.stack(cam, 0).contiguous() # torch.Size([240, 16, 8])
image = torch.stack(image,
0).contiguous() # torch.Size([240, 1, 3, 256, 128])
cam = cam.view(30, 8, 16, -1) # 8,8,16,8
image = image.view(30, 8, 1, 3, 256, -1)
path = PATH_EVAL + "fenzhi{}".format(k)
mkdir_if_missing(path)
for i in range(cam.size(0)):
fig = plt.figure(figsize=(15, 15))
for j in range(cam.size(1)):
ax1 = plt.subplot(3, 8, j + 1)
ax1.axis('off')
plt.title('cam', fontsize=18)
plt.imshow(cam[i][j].detach().cpu().numpy(), alpha=0.6, cmap='jet')
ax3 = plt.subplot(3, 8, j + 17)
ax3.axis('off')
plt.title('cam+img', fontsize=18)
cam_ij = cam[i][j].unsqueeze(0)
cam_ij = cam_ij.unsqueeze(0)
images_ij = image[i][j]
heatmap, raw_image = visualize(images_ij, cam_ij)
heatmap = heatmap.squeeze().cpu().numpy().transpose(1, 2, 0)
plt.imshow(heatmap)
ax4 = plt.subplot(3, 8, j + 9)
ax4.axis('off')
plt.title('raw_image', fontsize=18)
raw_image = raw_image.squeeze().cpu().numpy().transpose(1, 2, 0)
plt.imshow(raw_image)
# fig.tight_layout()
fig.savefig(PATH_EVAL + "fenzhi{}/cambatch_{}.jpg".format(k, i))
def download(self):
if self._check_integrity():
print("Files already downloaded and verified")
return
raw_dir = osp.join(self.root, 'raw')
mkdir_if_missing(raw_dir)
fpath1 = osp.join(raw_dir, datasetname + '.tar')
fpath2 = osp.join(raw_dir, flowname + '.tar')
if osp.isfile(fpath1) and osp.isfile(fpath2):
print("Using the download file:" + fpath1 + " " + fpath2)
else:
print("Please firstly download the files")
raise RuntimeError("Downloaded file missing!")
def download(self):
if self._check_integrity():
print("Files already downloaded and verified")
return
raw_dir = osp.join(self.root, 'raw')
mkdir_if_missing(raw_dir)
fpath1 = osp.join(raw_dir, datasetname + '.tar')
fpath2 = osp.join(raw_dir, flowname + '.tar')
if osp.isfile(fpath1) and osp.isfile(fpath2):
print("Using the download file:" + fpath1 + " " + fpath2)
else:
print("Please firstly download the files")
raise RuntimeError("Downloaded file missing!")
def _cp_img_to(src, dst, rank, prefix):
"""
Args:
src: image path or tuple (for vidreid)
dst: target directory # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual/0016C1T0006F001.jpg'
rank: int, denoting ranked position, starting from 1
prefix: string （query or gallery）
"""
if isinstance(src, tuple) or isinstance(src, list): # video reid
dst = osp.join(
dst, prefix + '_top' + str(rank).zfill(3)
) # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual/0016C1T0006F001.jpg/query_top000'
mkdir_if_missing(dst)
for img_path in src: # 将图片copy到目标文件夹中
shutil.copy(img_path, dst)
else:
dst = osp.join(
dst, prefix + '_top' + str(rank).zfill(3) + '_name_' +
osp.basename(src))
shutil.copy(src, dst)
def visual_batch(cam, image, k, save_dir, mode):
cam = cam.squeeze(
) # b, t, 16, 8 torch.Size([240, 16, 8])
b, t, h, w = cam.size(
) # b, t, 16, 8 torch.Size([240, 16, 8])
image = torch.stack(image, 0).contiguous(
) # [bt, 1, 3, 256, 128] torch.Size([240, 1, 3, 256, 128])
# cam = cam.view(8, 8, *cam.size()[1:]) # 8,8,16,8
image = image.view(b, t, 1, *image.size()[-3:]) # b, t, 1, 3, 256, 128
path = PATH + save_dir
mkdir_if_missing(path)
for i in range(cam.size(0)): # b
fig = plt.figure(figsize=(15, 15))
for j in range(cam.size(1)): # t
ax1 = plt.subplot(3, 8, j + 1)
ax1.axis('off')
plt.title('cam', fontsize=18)
plt.imshow(cam[i][j].detach().cpu().numpy(), alpha=0.6, cmap='jet')
ax3 = plt.subplot(3, 8, j + 17)
ax3.axis('off')
plt.title('cam+img', fontsize=18)
cam_ij = cam[i][j].unsqueeze(0)
cam_ij = cam_ij.unsqueeze(0)
images_ij = image[i][j]
heatmap, raw_image = visualize(images_ij, cam_ij)
heatmap = heatmap.squeeze().cpu().numpy().transpose(1, 2, 0)
plt.imshow(heatmap)
ax4 = plt.subplot(3, 8, j + 9)
ax4.axis('off')
plt.title('raw_image', fontsize=18)
raw_image = raw_image.squeeze().cpu().numpy().transpose(1, 2, 0)
plt.imshow(raw_image)
# fig.tight_layout()
fig.savefig(path + "/iter_{}index_{}_{}.jpg".format(k, i, mode))
def imgextract(self):
raw_dir = osp.join(self.root, 'raw')
exdir1 = osp.join(raw_dir, datasetname)
exdir2 = osp.join(raw_dir, flowname)
fpath1 = osp.join(raw_dir, datasetname + '.tar')
fpath2 = osp.join(raw_dir, flowname + '.tar')
if not osp.isdir(exdir1):
print("Extracting tar file")
cwd = os.getcwd()
tar = tarfile.open(fpath1)
mkdir_if_missing(exdir1)
os.chdir(exdir1)
tar.extractall()
tar.close()
os.chdir(cwd)
if not osp.isdir(exdir2):
print("Extracting tar file")
cwd = os.getcwd()
tar = tarfile.open(fpath2)
mkdir_if_missing(exdir2)
os.chdir(exdir2)
tar.extractall()
tar.close()
os.chdir(cwd)
# reorganzing the dataset
# Format
temp_images_dir = osp.join(self.root, 'temp_images')
mkdir_if_missing(temp_images_dir)
temp_others_dir = osp.join(self.root, 'temp_others')
mkdir_if_missing(temp_others_dir)
images_dir = osp.join(self.root, 'images')
mkdir_if_missing(images_dir)
others_dir = osp.join(self.root, 'others')
mkdir_if_missing(others_dir)
fpaths1 = sorted(glob(osp.join(exdir1, 'i-LIDS-VID/sequences', '*/*/*.png')))
fpaths2 = sorted(glob(osp.join(exdir2, flowname, '*/*/*.png')))
identities_imgraw = [[[] for _ in range(2)] for _ in range(319)]
identities_otherraw = [[[] for _ in range(2)] for _ in range(319)]
# image information
for fpath in fpaths1:
fname = osp.basename(fpath)
fname_list = fname.split('_')
cam_name = fname_list[0]
pid_name = fname_list[1]
cam = int(cam_name[-1])
pid = int(pid_name[-3:])
temp_fname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid, cam, len(identities_imgraw[pid - 1][cam - 1])))
identities_imgraw[pid - 1][cam - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_images_dir, temp_fname))
identities_temp = [x for x in identities_imgraw if x != [[], []]]
identities_images = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_images[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid, cam, img))
identities_images[pid][cam][img] = fname
shutil.copy(osp.join(temp_images_dir, temp_fname), osp.join(images_dir, fname))
shutil.rmtree(temp_images_dir)
# flow information
for fpath in fpaths2:
fname = osp.basename(fpath)
fname_list = fname.split('_')
cam_name = fname_list[0]
pid_name = fname_list[1]
cam = int(cam_name[-1])
pid = int(pid_name[-3:])
temp_fname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid, cam, len(identities_otherraw[pid - 1][cam - 1])))
identities_otherraw[pid - 1][cam - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_others_dir, temp_fname))
identities_temp = [x for x in identities_otherraw if x != [[], []]]
identities_others = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_others[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid, cam, img))
identities_others[pid][cam][img] = fname
shutil.copy(osp.join(temp_others_dir, temp_fname), osp.join(others_dir, fname))
shutil.rmtree(temp_others_dir)
meta = {'name': 'iLIDS-sequence', 'shot': 'sequence', 'num_cameras': 2,
'identities': identities_images}
write_json(meta, osp.join(self.root, 'meta.json'))
# Consider fixed training and testing split
splitmat_name = osp.join(exdir1, 'train-test people splits', 'train_test_splits_ilidsvid.mat')
data = sio.loadmat(splitmat_name)
person_list = data['ls_set']
num = len(identities_images)
splits = []
for i in range(10):
pids = (person_list[i] - 1).tolist()
trainval_pids = sorted(pids[:num // 2])
test_pids = sorted(pids[num // 2:])
split = {'trainval': trainval_pids,
'query': test_pids,
'gallery': test_pids}
splits.append(split)
write_json(splits, osp.join(self.root, 'splits.json'))
def main():
global best_prec
global opt
if opt['id'] != '':
model_id = opt['id']
else:
model_id = time.strftime("%m_%d_%H-%M-%S")
sys.stdout = Logger(osp.join(opt['log_dir'], 'log.' + model_id + '.txt'))
# initialize
checkpoint_dir = osp.join(opt['checkpoint_dir'], model_id)
mkdir_if_missing(checkpoint_dir)
# check gpu
assert opt['gpus'] is not None
# set random seed
cudnn.benchmark = False
cudnn.deterministic = True
random.seed(opt['seed'])
np.random.seed(opt['seed'])
torch.manual_seed(opt['seed'])
torch.cuda.manual_seed_all(opt['seed'])
# load imdb
train_refdb = get_db('refvg_train_' + opt['model_method'])
vocab = train_refdb.load_dictionary()
opt['vocab_size'] = len(vocab)
val_refdb = get_db('refvg_val_' + opt['model_method'])
# model, criterion, optimizer
model = SGReason(opt)
model = torch.nn.DataParallel(model).cuda()
criterion = SoftmaxLoss().cuda()
optimizer = torch.optim.Adam(list(model.parameters()) +
list(criterion.parameters()),
lr=opt['learning_rate'],
betas=(opt['optim_alpha'], opt['optim_beta']),
eps=opt['optim_epsilon'])
scheduler = ReduceLROnPlateau(optimizer,
factor=0.1,
patience=3,
mode='max')
if opt['evaluate']:
if osp.isfile(opt['model']):
model, criterion = load_checkpoint(model, criterion, opt['model'])
test_refdb = get_db('refvg_test_' + opt['model_method'])
test_dataset = RefDataset(test_refdb, vocab, opt)
test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=opt['batch_size'],
shuffle=False,
num_workers=opt['workers'],
pin_memory=True)
test_loss, test_prec = validate(test_loader, model, criterion)
print(test_loss, test_prec)
else:
print("=> no checkpoint found at '{}'".format(opt['model']))
return
# start training
epoch_cur = 0
train_dataset = RefDataset(train_refdb, vocab, opt)
val_dataset = RefDataset(val_refdb, vocab, opt)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=opt['batch_size'],
shuffle=True,
num_workers=opt['workers'],
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=opt['batch_size'],
shuffle=False,
num_workers=opt['workers'],
pin_memory=True)
for epoch in range(epoch_cur, opt['max_epochs']):
train(train_loader, model, criterion, optimizer, epoch)
val_loss, prec = validate(val_loader, model, criterion, epoch)
scheduler.step(prec)
for i, param_group in enumerate(optimizer.param_groups):
print(float(param_group['lr']))
is_best = prec >= best_prec
best_prec = max(best_prec, prec)
save_checkpoint(
{
'model_state_dict': model.state_dict(),
'crit_state_dict': criterion.state_dict(),
'optimizer': optimizer.state_dict()
}, is_best, checkpoint_dir, str(epoch))
def visualize_ranked_results(distmat,
queryloader,
galleryloader,
save_dir='',
visual_id=2,
topk=10):
"""Visualizes ranked results. 存放在一个文件夹中
Supports both image-reid and video-reid.
Args:
distmat (numpy.ndarray): distance matrix of shape (num_query, num_gallery).
queryloader (tuple): tuples of (img_path(s), pid, camid).
galleryloader (tuple): tuples of (img_path(s), pid, camid).
save_dir (str): directory to save output images.
visual_id(int, optional): only show 1 id
topk (int, optional): denoting top-k images in the rank list to be visualized.
"""
num_q, num_g = distmat.shape
print('Visualizing top-{} ranks'.format(topk))
print('# query: {}\n# gallery {}'.format(num_q, num_g))
print('Saving images to "{}"'.format(save_dir))
query = queryloader # 1980个tuple (img_path(s), pid, camid)）
gallery = galleryloader # 9330个tuple (img_path(s), pid, camid)
assert num_q == len(query)
assert num_g == len(gallery)
indices = np.argsort(distmat, axis=1) # <class 'tuple'>: (1980, 9330)
mkdir_if_missing(
save_dir
) # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual'
def _cp_img_to(src, dst, rank, prefix):
"""
Args:
src: image path or tuple (for vidreid)
dst: target directory # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual/0016C1T0006F001.jpg'
rank: int, denoting ranked position, starting from 1
prefix: string （query or gallery）
"""
if isinstance(src, tuple) or isinstance(src, list): # video reid
dst = osp.join(
dst, prefix + '_top' + str(rank).zfill(3)
) # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual/0016C1T0006F001.jpg/query_top000'
mkdir_if_missing(dst)
for img_path in src: # 将图片copy到目标文件夹中
shutil.copy(img_path, dst)
else:
dst = osp.join(
dst, prefix + '_top' + str(rank).zfill(3) + '_name_' +
osp.basename(src))
shutil.copy(src, dst)
for q_idx in range(
num_q): # 考虑到速度等因素，只输出1个id的rank结果。这个id不是实际的行人id，是在tuple中的顺序
if q_idx == visual_id: # 14
qimg_path, qpid, qcamid = query[q_idx] # qpid = 16， camid = 0
if isinstance(qimg_path, tuple) or isinstance(
qimg_path,
list): # query_dir 保存Rank结果的文件夹名称 = query的第一张图片名称
qdir = osp.join(
save_dir, osp.basename(qimg_path[0])
) # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual/0016C1T0006F001.jpg'
else:
qdir = osp.join(save_dir, osp.basename(qimg_path))
mkdir_if_missing(qdir) # 新建这个保存rank结果的文件夹
_cp_img_to(qimg_path, qdir, rank=0,
prefix='query') # 复制query的图片到结果文件夹中
rank_idx = 1
for g_idx in indices[q_idx, :]: # 3291， 3288， 3289， 3290， 3293
gimg_path, gpid, gcamid = gallery[g_idx]
invalid = (qpid == gpid) & (qcamid == gcamid
) # true， 排除相同cam的情况
if not invalid:
_cp_img_to(gimg_path,
qdir,
rank=rank_idx,
prefix='gallery')
rank_idx += 1
if rank_idx > topk:
break
print("Done")
def visualize_in_pic(distmat,
queryloader,
galleryloader,
save_dir='',
visual_id=2,
topk=9):
"""
distmat (numpy.ndarray): distance matrix of shape (num_query, num_gallery).
queryloader (tuple): tuples of (img_path(s), pid, camid).
galleryloader (tuple): tuples of (img_path(s), pid, camid).
save_dir (str): directory to save output images.
visual_id(int, optional): only show 1 id
topk (int, optional): denoting top-k images in the rank list to be visualized.
"""
num_q, num_g = distmat.shape
print('Visualizing top-{} ranks'.format(topk + 1))
print('# query: {}\n# gallery {}'.format(num_q, num_g))
print('Saving images to "{}"'.format(save_dir))
query = queryloader # 1980个tuple (img_path(s), pid, camid)）
gallery = galleryloader # 9330个tuple (img_path(s), pid, camid)
assert num_q == len(query)
assert num_g == len(gallery)
indices = np.argsort(distmat, axis=1) # <class 'tuple'>: (1980, 9330)
mkdir_if_missing(
save_dir
) # '/home/ying/Desktop/mars_rank/log/debug_for_eval/split0visual'
def imshow(path, title=None):
"""Imshow for Tensor."""
im = plt.imread(path)
plt.imshow(im)
if title is not None:
plt.title(title)
# plt.pause(0.001) # pause a bit so that plots are updated
flag = 0
for q_idx in range(num_q): # 考虑到速度等因素，只输出1个id的rank结果。2,4,6,8,10..
qimg_path, qpid, qcamid = query[q_idx] # qpid = 16， camid = 0
if qpid == visual_id: # 14
flag = 1
fig = plt.figure(figsize=(25, 8))
ax = plt.subplot(1, 11, 1)
ax.axis('off')
imshow(qimg_path[0], 'query, pid:{}'.format(qpid))
rank_idx = 0
for g_idx in indices[q_idx, :]: # 3291， 3288， 3289， 3290， 3293
gimg_path, gpid, gcamid = gallery[g_idx]
# invalid = (qpid == gpid) & (qcamid == gcamid) # true， 排除相同cam的情况
invalid = False
if not invalid:
rank_idx += 1
ax = plt.subplot(1, 11, rank_idx + 1)
ax.axis('off')
imshow(gimg_path[0])
if qpid == gpid:
ax.set_title('rank:{},pid{}_{}'.format(
rank_idx, gpid, gcamid),
color='green')
else:
ax.set_title('rank:{},pid{}_{}'.format(
rank_idx, gpid, gcamid),
color='red')
if rank_idx > topk:
break
fig.savefig("show_{}_{}.png".format(qpid, qcamid))
break
if flag == 1:
print("Done")
else:
print("No matched person in query_dataset, try another id")
def imgextract(self):
raw_dir = osp.join(self.root, 'raw')
# raw_dir = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw
exdir1 = osp.join(raw_dir, datasetname)
# exdir1 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid_2011
exdir2 = osp.join(raw_dir, flowname)
# exdir2 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid2011flow
fpath1 = osp.join(raw_dir, datasetname + '.tar')
# fpath1 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid_2011.tar
fpath2 = osp.join(raw_dir, flowname + '.tar')
# fpath2 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid2011flow.tar
if not osp.isdir(exdir1):
print("Extracting tar file")
cwd = os.getcwd()
tar_ref = tarfile.open(fpath1)
mkdir_if_missing(exdir1)
os.chdir(exdir1)
tar_ref.extractall()
tar_ref.close()
os.chdir(cwd)
if not osp.isdir(exdir2):
print("Extracting tar file")
cwd = os.getcwd()
tar_ref = tarfile.open(fpath2)
mkdir_if_missing(exdir2)
os.chdir(exdir2)
tar_ref.extractall()
tar_ref.close()
os.chdir(cwd)
# recognizing the dataset
# Format
temp_images_dir = osp.join(self.root, 'temp_images')
mkdir_if_missing(temp_images_dir)
temp_others_dir = osp.join(self.root, 'temp_others')
mkdir_if_missing(temp_others_dir)
images_dir = osp.join(self.root, 'images')
mkdir_if_missing(images_dir)
# images_dir = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/images
others_dir = osp.join(self.root, 'others')
mkdir_if_missing(others_dir)
# others_dir = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/others
fpaths1 = sorted(glob(osp.join(exdir1, 'prid_2011/multi_shot', '*/*/*.png')))
fpaths2 = sorted(glob(osp.join(exdir2, 'prid2011flow', '*/*/*.png')))
identities_imgraw = [[[] for _ in range(2)] for _ in range(200)]
identities_otherraw = [[[] for _ in range(2)] for _ in range(200)]
for fpath in fpaths1:
fname = fpath
fname_list = fname.split('/')
cam_name = fname_list[-3] # cam_a / cam_b
pid_name = fname_list[-2] # person_001
frame_name = fname_list[-1] # 0001.png
cam_id = 1 if cam_name == 'cam_a' else 2 # cam_id = 1 / 2
pid_id = int(pid_name.split('_')[-1]) # pid_id = 001
if pid_id > 200:
continue
frame_id = int(frame_name.split('.')[-2]) # frame_id = 0001
temp_fname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid_id-1, cam_id-1, frame_id-1))
identities_imgraw[pid_id - 1][cam_id - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_images_dir, temp_fname))
identities_temp = [x for x in identities_imgraw if x != [[], []]]
identities_images = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_images[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'.format(pid, cam, img))
identities_images[pid][cam][img] = fname
shutil.copy(osp.join(temp_images_dir, temp_fname), osp.join(images_dir, fname))
shutil.rmtree(temp_images_dir)
for fpath in fpaths2:
fname = fpath
fname_list = fname.split('/')
cam_name = fname_list[-3] # cam_a / cam_b
pid_name = fname_list[-2] # person_001
frame_name = fname_list[-1] # 0001.png
cam_id = 1 if cam_name == 'cam_a' else 2 # cam_id = 1 / 2
pid_id = int(pid_name.split('_')[-1]) # pid_id = 001
if pid_id > 200:
continue
frame_id = int(frame_name.split('.')[-2]) # frame_id = 0001
temp_fname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid_id-1, cam_id-1, frame_id-1))
identities_otherraw[pid_id - 1][cam_id - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_others_dir, temp_fname))
identities_temp = [x for x in identities_otherraw if x != [[], []]]
identities_others = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_others[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'.format(pid, cam, img))
identities_images[pid][cam][img] = fname
shutil.copy(osp.join(temp_others_dir, temp_fname), osp.join(others_dir, fname))
shutil.rmtree(temp_others_dir)
meta = {'name': 'prid-sequence', 'shot': 'sequence', 'num_cameras': 2,
'identities': identities_images}
write_json(meta, osp.join(self.root, 'meta.json'))
# Consider fixed training and testing split
num = 200
splits = []
for i in range(10):
pids = np.random.permutation(num)
pids = (pids - 1).tolist()
trainval_pids = pids[:num // 2]
test_pids = pids[num // 2:]
split = {'trainval': trainval_pids,
'query': test_pids,
'gallery': test_pids}
splits.append(split)
write_json(splits, osp.join(self.root, 'splits.json'))
def imgextract(self):
raw_dir = osp.join(self.root, 'raw')
# raw_dir = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw
exdir1 = osp.join(raw_dir, datasetname)
# exdir1 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid_2011
exdir2 = osp.join(raw_dir, flowname)
# exdir2 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid2011flow
fpath1 = osp.join(raw_dir, datasetname + '.tar')
# fpath1 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid_2011.tar
fpath2 = osp.join(raw_dir, flowname + '.tar')
# fpath2 = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/raw/prid2011flow.tar
if not osp.isdir(exdir1):
print("Extracting tar file")
cwd = os.getcwd()
tar_ref = tarfile.open(fpath1)
mkdir_if_missing(exdir1)
os.chdir(exdir1)
tar_ref.extractall()
tar_ref.close()
os.chdir(cwd)
if not osp.isdir(exdir2):
print("Extracting tar file")
cwd = os.getcwd()
tar_ref = tarfile.open(fpath2)
mkdir_if_missing(exdir2)
os.chdir(exdir2)
tar_ref.extractall()
tar_ref.close()
os.chdir(cwd)
# recognizing the dataset
# Format
temp_images_dir = osp.join(self.root, 'temp_images')
mkdir_if_missing(temp_images_dir)
temp_others_dir = osp.join(self.root, 'temp_others')
mkdir_if_missing(temp_others_dir)
images_dir = osp.join(self.root, 'images')
mkdir_if_missing(images_dir)
# images_dir = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/images
others_dir = osp.join(self.root, 'others')
mkdir_if_missing(others_dir)
# others_dir = /media/ying/0BDD17830BDD1783/video_reid _prid/data/prid2011sequence/others
fpaths1 = sorted(
glob(osp.join(exdir1, 'prid_2011/multi_shot',
'*/*/*.png'))) # 存放所有图片的绝对路径
fpaths2 = sorted(glob(osp.join(exdir2, 'prid2011flow', '*/*/*.png')))
identities_imgraw = [[[] for _ in range(2)]
for _ in range(200)] # 200个[ []..[] ]
identities_otherraw = [[[] for _ in range(2)] for _ in range(200)]
for fpath in fpaths1:
fname = fpath
fname_list = fname.split('/')
cam_name = fname_list[-3] # cam_a / cam_b
pid_name = fname_list[-2] # person_001
frame_name = fname_list[-1] # 0001.png
cam_id = 1 if cam_name == 'cam_a' else 2 # cam_id = 1 / 2
pid_id = int(pid_name.split('_')[-1]) # pid_id = 001
if pid_id > 200:
continue
frame_id = int(frame_name.split('.')[-2]) # frame_id = 0001
temp_fname = ('{:08d}_{:02d}_{:04d}.png'.format(
pid_id - 1, cam_id - 1, frame_id - 1))
identities_imgraw[pid_id - 1][cam_id - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_images_dir, temp_fname))
identities_temp = [x for x in identities_imgraw if x != [[], []]]
identities_images = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_images[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'.format(pid, cam, img))
identities_images[pid][cam][img] = fname
shutil.copy(osp.join(temp_images_dir, temp_fname),
osp.join(images_dir, fname))
shutil.rmtree(temp_images_dir)
for fpath in fpaths2:
fname = fpath
fname_list = fname.split('/')
cam_name = fname_list[-3] # cam_a / cam_b
pid_name = fname_list[-2] # person_001
frame_name = fname_list[-1] # 0001.png
cam_id = 1 if cam_name == 'cam_a' else 2 # cam_id = 1 / 2
pid_id = int(pid_name.split('_')[-1]) # pid_id = 001
if pid_id > 200:
continue
frame_id = int(frame_name.split('.')[-2]) # frame_id = 0001
temp_fname = ('{:08d}_{:02d}_{:04d}.png'.format(
pid_id - 1, cam_id - 1, frame_id - 1))
identities_otherraw[pid_id - 1][cam_id - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_others_dir, temp_fname))
identities_temp = [x for x in identities_otherraw if x != [[], []]]
identities_others = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_others[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'.format(pid, cam, img))
identities_images[pid][cam][img] = fname
shutil.copy(osp.join(temp_others_dir, temp_fname),
osp.join(others_dir, fname))
shutil.rmtree(temp_others_dir)
meta = {
'name': 'prid-sequence',
'shot': 'sequence',
'num_cameras': 2,
'identities': identities_images
}
write_json(meta, osp.join(self.root, 'meta.json'))
# Consider fixed training and testing split
num = len(identities_images)
splits = []
for i in range(20):
pids = np.random.permutation(num)
pids = (pids - 1).tolist()
trainval_pids = pids[:num // 2]
test_pids = pids[num // 2:]
split = {
'trainval': trainval_pids,
'query': test_pids,
'gallery': test_pids
}
splits.append(split)
write_json(splits, osp.join(self.root, 'splits.json'))
def imgextract(self):
raw_dir = osp.join(self.root, 'raw')
exdir1 = osp.join(raw_dir, datasetname)
exdir2 = osp.join(raw_dir, flowname)
fpath1 = osp.join(raw_dir, datasetname + '.tar')
fpath2 = osp.join(raw_dir, flowname + '.tar')
if not osp.isdir(exdir1):
print("Extracting tar file")
cwd = os.getcwd()
tar = tarfile.open(fpath1)
mkdir_if_missing(exdir1)
os.chdir(exdir1)
tar.extractall()
tar.close()
os.chdir(cwd)
if not osp.isdir(exdir2):
print("Extracting tar file")
cwd = os.getcwd()
tar = tarfile.open(fpath2)
mkdir_if_missing(exdir2)
os.chdir(exdir2)
tar.extractall()
tar.close()
os.chdir(cwd)
# reorganzing the dataset
# Format
temp_images_dir = osp.join(self.root, 'temp_images')
mkdir_if_missing(temp_images_dir)
temp_others_dir = osp.join(self.root, 'temp_others')
mkdir_if_missing(temp_others_dir)
images_dir = osp.join(self.root, 'images')
mkdir_if_missing(images_dir)
others_dir = osp.join(self.root, 'others')
mkdir_if_missing(others_dir)
fpaths1 = sorted(
glob(osp.join(exdir1, 'i-LIDS-VID/sequences', '*/*/*.png')))
fpaths2 = sorted(glob(osp.join(exdir2, flowname, '*/*/*.png')))
identities_imgraw = [[[] for _ in range(2)] for _ in range(319)]
identities_otherraw = [[[] for _ in range(2)] for _ in range(319)]
# image information
for fpath in fpaths1:
fname = osp.basename(fpath)
fname_list = fname.split('_')
cam_name = fname_list[0]
pid_name = fname_list[1]
cam = int(cam_name[-1])
pid = int(pid_name[-3:])
temp_fname = ('{:08d}_{:02d}_{:04d}.png'.format(
pid, cam, len(identities_imgraw[pid - 1][cam - 1])))
identities_imgraw[pid - 1][cam - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_images_dir, temp_fname))
identities_temp = [x for x in identities_imgraw if x != [[], []]]
identities_images = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_images[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'.format(
pid, cam, img))
identities_images[pid][cam][img] = fname
shutil.copy(osp.join(temp_images_dir, temp_fname),
osp.join(images_dir, fname))
shutil.rmtree(temp_images_dir)
# flow information
for fpath in fpaths2:
fname = osp.basename(fpath)
fname_list = fname.split('_')
cam_name = fname_list[0]
pid_name = fname_list[1]
cam = int(cam_name[-1])
pid = int(pid_name[-3:])
temp_fname = ('{:08d}_{:02d}_{:04d}.png'.format(
pid, cam, len(identities_otherraw[pid - 1][cam - 1])))
identities_otherraw[pid - 1][cam - 1].append(temp_fname)
shutil.copy(fpath, osp.join(temp_others_dir, temp_fname))
identities_temp = [x for x in identities_otherraw if x != [[], []]]
identities_others = identities_temp
for pid in range(len(identities_temp)):
for cam in range(2):
for img in range(len(identities_others[pid][cam])):
temp_fname = identities_temp[pid][cam][img]
fname = ('{:08d}_{:02d}_{:04d}.png'.format(
pid, cam, img))
identities_others[pid][cam][img] = fname
shutil.copy(osp.join(temp_others_dir, temp_fname),
osp.join(others_dir, fname))
shutil.rmtree(temp_others_dir)
meta = {
'name': 'iLIDS-sequence',
'shot': 'sequence',
'num_cameras': 2,
'identities': identities_images
}
write_json(meta, osp.join(self.root, 'meta.json'))
# Consider fixed training and testing split
splitmat_name = osp.join(exdir1, 'train-test people splits',
'train_test_splits_ilidsvid.mat')
data = sio.loadmat(splitmat_name)
person_list = data['ls_set']
num = len(identities_images)
splits = []
for i in range(10):
pids = (person_list[i] - 1).tolist()
trainval_pids = sorted(pids[:num // 2])
test_pids = sorted(pids[num // 2:])
split = {
'trainval': trainval_pids,
'query': test_pids,
'gallery': test_pids
}
splits.append(split)
write_json(splits, osp.join(self.root, 'splits.json'))
def imgextract(self):
raw_dir = osp.join(self.root, 'raw')
exdir1 = osp.join(raw_dir, datasetname)
exdir2 = osp.join(raw_dir, flowname)
fpath1 = osp.join(raw_dir, datasetname + '.zip')
fpath2 = osp.join(raw_dir, flowname + '.tar')
if not osp.isdir(exdir1):
print("Extracting tar file")
cwd = os.getcwd()
zip_ref = zipfile.ZipFile(fpath1, 'r')
mkdir_if_missing(exdir1)
zip_ref.extractall(exdir1)
zip_ref.close()
os.chdir(cwd)
if not osp.isdir(exdir2):
print("Extracting tar file")
cwd = os.getcwd()
tar_ref = tarfile.open(fpath2)
mkdir_if_missing(exdir2)
os.chdir(exdir2)
tar_ref.extractall()
tar_ref.close()
os.chdir(cwd)
## recognizing the dataset
# Format
images_dir = osp.join(self.root, 'images')
mkdir_if_missing(images_dir)
others_dir = osp.join(self.root, 'others')
mkdir_if_missing(others_dir)
fpaths1 = sorted(glob(osp.join(exdir1, 'multi_shot', '*/*/*.png')))
fpaths2 = sorted(glob(osp.join(exdir2, '*/*.png')))
identities_images = [[[] for _ in range(2)] for _ in range(200)]
for fpath in fpaths1:
fname = fpath
fname_list = fname.split('/')
cam_name = fname_list[-3]
pid_name = fname_list[-2]
frame_name = fname_list[-1]
cam_id = 1 if cam_name =='cam_a' else 2
pid_id = int(pid_name.split('_')[-1])
if pid_id > 200:
continue
frame_id = int(frame_name.split('.')[-2])
imagefname = ('{:08d}_{:02d}_{:04d}.png'
.format(pid_id-1, cam_id-1, frame_id-1))
identities_images[pid_id - 1][cam_id - 1].append(imagefname)
shutil.copy(fpath, osp.join(images_dir, imagefname))
for fpath in fpaths2:
fname = fpath
fname_list = fname.split('/')
fname_img = fname_list[-1]
shutil.copy(fname, osp.join(others_dir, fname_img))
meta = {'name': 'iLIDS-sequence', 'shot': 'sequence', 'num_cameras': 2,
'identities': identities_images}
write_json(meta, osp.join(self.root, 'meta.json'))
# Consider fixed training and testing split
num = 200
splits = []
for i in range(10):
pids = np.random.permutation(num)
pids = (pids -1).tolist()
trainval_pids = pids[:num // 2]
test_pids = pids[num // 2:]
split = {'trainval': trainval_pids,
'query': test_pids,
'gallery': test_pids}
splits.append(split)
write_json(splits, osp.join(self.root, 'splits.json'))
def main(args):
if not torch.cuda.is_available():
args.cpu_only = True
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if not args.cpu_only:
torch.cuda.manual_seed_all(args.seed)
cudnn.benchmark = True
# Logs directory
mkdir_if_missing(args.logs_dir)
if not args.eval_only:
sys.stdout = Logger(osp.join(args.logs_dir, 'log.txt'))
# Data
train_dataset, test_dataset, num_classes = get_datasets(
args.dataset, args.data_dir)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=args.workers,
shuffle=True,
pin_memory=True)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
num_workers=args.workers,
shuffle=False,
pin_memory=True)
# Model
model = WideResNet(args.depth,
args.width,
num_classes,
dropout_rate=args.dropout)
criterion = nn.CrossEntropyLoss()
start_epoch, best_prec1 = 0, 0
if args.resume:
checkpoint = load_checkpoint(args.resume)
model.load_state_dict(checkpoint['model'])
start_epoch = checkpoint['epoch'] + 1
best_prec1 = checkpoint['best_prec1']
print("=> Load from {}, start epoch {}, best prec1 {:.2%}".format(
args.resume, start_epoch, best_prec1))
if not args.cpu_only:
model = DataParallel(model).cuda()
criterion = criterion.cuda()
# Optimizer
if args.optim_method == 'sgd':
optimizer = SGD(model.parameters(),
lr=args.lr,
nesterov=True,
momentum=0.9,
weight_decay=args.weight_decay)
else:
optimizer = Adam(model.parameters(), lr=args.lr)
# Evaluation only
if args.eval_only:
evaluate(start_epoch - 1, test_loader, model, criterion, args.cpu_only)
return
# Training
epoch_steps = json.loads(args.epoch_steps)[::-1]
for epoch in range(start_epoch, args.epochs):
# Adjust learning rate
power = 0
for i, step in enumerate(epoch_steps):
if epoch >= step:
power = len(epoch_steps) - i
lr = args.lr * (args.lr_decay_ratio**power)
for g in optimizer.param_groups:
g['lr'] = lr
# Training
train(epoch, train_loader, model, criterion, optimizer, args.cpu_only)
prec1 = evaluate(epoch, test_loader, model, criterion, args.cpu_only)
is_best = prec1 > best_prec1
best_prec1 = max(best_prec1, prec1)
# Save checkpoint
checkpoint = {'epoch': epoch, 'best_prec1': best_prec1}
if args.cpu_only:
checkpoint['model'] = model.state_dict()
else:
checkpoint['model'] = model.module.state_dict()
save_checkpoint(checkpoint, is_best,
osp.join(args.logs_dir, 'checkpoint.pth.tar'))
print('\n * Finished epoch {} Prec1: {:.2%} Best: {:.2%}{}\n'.format(
epoch, prec1, best_prec1, ' *' if is_best else ''))
diff = np.absolute(img2 - img1)
# diff = cv2.cvtColor(diff, cv2.COLOR_BGR2GRAY)
diff = cv2.normalize(diff, None, mmin, mmax, norm_type=cv2.NORM_MINMAX)
diff = cv2.resize(diff, (240, 240))
diff = np.stack((diff, ) * 3, axis=-1)
cv2.imwrite(diff_path + '/diff.png', diff)
return is_align
if __name__ == '__main__':
working_dir = osp.dirname(osp.abspath(__file__))
data_dir = "data"
dataset = 'GTOS_256'
dataset_dir = osp.join(data_dir, dataset)
img_dir = osp.join(dataset_dir, 'images')
diff_dir = osp.join(dataset_dir, 'diff')
mkdir_if_missing(diff_dir)
cnt = 0
for idx, scene in enumerate(sorted(os.listdir(img_dir))):
img_scene_dir = osp.join(img_dir, scene)
if not osp.isdir(img_scene_dir):
continue
diff_scene_dir = osp.join(diff_dir, scene)
mkdir_if_missing(diff_scene_dir)
if not make_diff(img_scene_dir, diff_scene_dir):
cnt += 1
print(cnt, " images cannot be successfully aligned")
