def read(filenames, max_d, interval_scale):
ref_name, ref_cam_name, srcs_name, srcs_cam_name, gt_name, masks_name, skip = [
filenames[attr] for attr in
['ref', 'ref_cam', 'srcs', 'srcs_cam', 'gt', 'masks', 'skip']
]
ref, *srcs = [cv2.imread(fn) for fn in [ref_name] + srcs_name]
ref_cam, *srcs_cam = [
load_cam(fn, max_d, interval_scale)
for fn in [ref_cam_name] + srcs_cam_name
]
gt = np.expand_dims(load_pfm(gt_name), -1)
masks = [
np.expand_dims(cv2.imread(fn, cv2.IMREAD_GRAYSCALE), -1)
for fn in masks_name
]
# masks = [(np.ones_like(gt)*255).astype(np.uint8) for fn in masks_name]
return {
'ref': ref,
'ref_cam': ref_cam,
'srcs': srcs,
'srcs_cam': srcs_cam,
'gt': gt,
'masks': masks,
'skip': skip
}
def read(filenames):
ref_name, ref_cam_name, srcs_name, srcs_cam_name, ref_depth_name, srcs_depth_name, ref_probs_name = [
filenames[attr] for attr in [
'ref', 'ref_cam', 'srcs', 'srcs_cam', 'ref_depth', 'srcs_depth',
'ref_probs'
]
]
ref, *srcs = [cv2.imread(fn) for fn in [ref_name] + srcs_name]
ref_cam, *srcs_cam = [
load_cam(fn, 0, 1) for fn in [ref_cam_name] + srcs_cam_name
]
ref_depth, *srcs_depth = [
np.expand_dims(load_pfm(fn), axis=-1)
for fn in [ref_depth_name] + srcs_depth_name
]
ref_probs = [
np.expand_dims(load_pfm(fn), axis=-1) for fn in ref_probs_name
]
return {
'ref': ref,
'ref_cam': ref_cam,
'ref_depth': ref_depth,
'srcs': srcs,
'srcs_cam': srcs_cam,
'srcs_depth': srcs_depth,
'ref_probs': ref_probs,
'skip': filenames['skip'],
'id': filenames['id']
}
def read(filenames, max_d, interval_scale):
ref_name, ref_cam_name, srcs_name, srcs_cam_name, skip = [filenames[attr] for attr in ['ref', 'ref_cam', 'srcs', 'srcs_cam', 'skip']]
ref, *srcs = [cv2.imread(fn) for fn in [ref_name] + srcs_name]
ref_cam, *srcs_cam = [load_cam(fn, max_d, interval_scale) for fn in [ref_cam_name] + srcs_cam_name]
gt = np.zeros((ref.shape[0], ref.shape[1], 1))
masks = [np.zeros((ref.shape[0], ref.shape[1], 1)) for i in range(len(srcs))]
return {
'ref': ref,
'ref_cam': ref_cam,
'srcs': srcs,
'srcs_cam': srcs_cam,
'gt': gt,
'masks': masks,
'skip': skip
}
def read(filenames, num_src):
ref_name, ref_cam_name, srcs_name, srcs_cam_name, gt_name = [
filenames[attr]
for attr in ['ref', 'ref_cam', 'srcs', 'srcs_cam', 'gt']
]
ref, *srcs = [cv2.imread(fn) for fn in [ref_name] + srcs_name[:num_src]]
ref_cam, *srcs_cam = [
load_cam(fn) for fn in [ref_cam_name] + srcs_cam_name[:num_src]
]
gt = np.expand_dims(load_pfm(gt_name), -1)
masks = [(np.ones_like(gt) * 255).astype(np.uint8)
for _ in range(len(srcs))]
return {
'ref': ref,
'ref_cam': ref_cam,
'srcs': srcs,
'srcs_cam': srcs_cam,
'gt': gt,
'masks': masks,
'skip': 0
}
def read(filenames, max_d, interval_scale):
ref_name, ref_cam_name, srcs_name, srcs_cam_name, gt_name, skip = [
filenames[attr]
for attr in ['ref', 'ref_cam', 'srcs', 'srcs_cam', 'gt', 'skip']
]
ref, *srcs = [
cv2.imread(fn) if fn != 'dummy' else None
for fn in [ref_name] + srcs_name
]
srcs = [
src if src is not None else np.ones_like(ref, dtype=np.uint8)
for src in srcs
]
ref_cam, *srcs_cam = [
load_cam(fn, max_d, interval_scale) if fn != 'dummy' else None
for fn in [ref_cam_name] + srcs_cam_name
]
srcs_cam = [
src_cam if src_cam is not None else np.ones_like(ref_cam,
dtype=np.float32)
for src_cam in srcs_cam
]
gt = np.expand_dims(load_pfm(gt_name), -1)
masks = [(np.ones_like(gt) * 255).astype(np.uint8)
for _ in range(len(srcs))]
if ref_cam[1, 3, 0] <= 0:
skip = 1
print(f'depth start <= 0')
return {
'ref': ref,
'ref_cam': ref_cam,
'srcs': srcs,
'srcs_cam': srcs_cam,
'gt': gt,
'masks': masks,
'skip': skip
}
parser.add_argument('--cam_scale', type=float, default=1)
# parser.add_argument('--show_result', action='store_true', default=False)
parser.add_argument('--downsample', type=float, default=None)
args = parser.parse_args()
pthresh = [float(v) for v in args.pthresh.split(',')]
num_src = args.view
pair = load_pair(args.pair, min_views=num_src)
n_views = len(pair['id_list'])
views = {}
for i, id in tqdm(enumerate(pair['id_list']), 'load data', n_views):
image = cv2.imread(f'{args.data}/{id.zfill(8)}.jpg').transpose(2,0,1)[::-1]
cam = load_cam(f'{args.data}/cam_{id.zfill(8)}_flow3.txt', 256, 1)
depth = np.expand_dims(load_pfm(f'{args.data}/{id.zfill(8)}_flow3.pfm'), axis=0)
probs = np.stack([load_pfm(f'{args.data}/{id.zfill(8)}_flow{k+1}_prob.pfm') for k in range(3)], axis=0)
views[id] = {
'image': image, # 13hw (after next step)
'cam': cam, # 1244
'depth': depth, # 11hw
'prob': probs, # 13hw
}
recursive_apply(views[id], lambda arr: torch.from_numpy(np.ascontiguousarray(arr)).float().unsqueeze(0))
for i, id in tqdm(enumerate(pair['id_list']), 'prob filter', n_views):
views[id]['mask'] = prob_filter(views[id]['prob'].cuda(), pthresh).cpu() # 11hw bool
views[id]['depth'] *= views[id]['mask']
update = {}
