def estimateMidas(img, msize):
transform = Compose([
Resize(
msize,
msize,
resize_target=None,
keep_aspect_ratio=True,
ensure_multiple_of=32,
resize_method="upper_bound",
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
PrepareForNet(),
])
img_input = transform({"image": img})["image"]
# compute
with torch.no_grad():
sample = torch.from_numpy(img_input).to(device).unsqueeze(0)
prediction = midasmodel.forward(sample)
torch.cuda.empty_cache()
prediction = prediction.squeeze().cpu().numpy()
prediction = cv2.resize(prediction, (img.shape[1], img.shape[0]),
interpolation=cv2.INTER_CUBIC)
depth_min = prediction.min()
depth_max = prediction.max()
if depth_max - depth_min > numpy.finfo("float").eps:
prediction = (prediction - depth_min) / (depth_max - depth_min)
else:
prediction = 0
return prediction
def estimatemidas(img, msize):
# MiDas -v2 forward pass script adapted from https://github.com/intel-isl/MiDaS/tree/v2
transform = Compose(
[
Resize(
msize,
msize,
resize_target=None,
keep_aspect_ratio=True,
ensure_multiple_of=32,
resize_method="upper_bound",
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
PrepareForNet(),
]
)
img_input = transform({"image": img})["image"]
# Forward pass
with torch.no_grad():
sample = torch.from_numpy(img_input).to(device).unsqueeze(0)
prediction = midasmodel.forward(sample)
prediction = prediction.squeeze().cpu().numpy()
prediction = cv2.resize(prediction, (img.shape[1], img.shape[0]), interpolation=cv2.INTER_CUBIC)
# Normalization
depth_min = prediction.min()
depth_max = prediction.max()
if depth_max - depth_min > np.finfo("float").eps:
prediction = (prediction - depth_min) / (depth_max - depth_min)
else:
prediction = 0
return prediction