The essence of creating dense depth map is to calculate weighted depth information according to the distance of neighborhood. The original repo is based on inverse distance weighted. When it comes to the distance calculation, I have mads some adjustments, replacing the np.round()
with np.int32()
. Also, in the double loop statements, I have deleted - grid - 1
.
`
def dense_map(Pts, n, m, grid):
ng = 2 * grid + 1
mX = np.zeros((m,n)) + np.float("inf")
mY = np.zeros((m,n)) + np.float("inf")
mD = np.zeros((m,n))
mX[np.int32(Pts[1]),np.int32(Pts[0])] = Pts[0] - np.int32(Pts[0])#np.round(Pts[0]) # NOTICE
mY[np.int32(Pts[1]),np.int32(Pts[0])] = Pts[1] - np.int32(Pts[1])#np.round(Pts[1]) # NOTICE
mD[np.int32(Pts[1]),np.int32(Pts[0])] = Pts[2]
KmX = np.zeros((ng, ng, m - ng, n - ng))
KmY = np.zeros((ng, ng, m - ng, n - ng))
KmD = np.zeros((ng, ng, m - ng, n - ng))
for i in range(ng):
for j in range(ng):
KmX[i,j] = mX[i : (m - ng + i), j : (n - ng + j)] + i #- grid - 1 # NOTICE
KmY[i,j] = mY[i : (m - ng + i), j : (n - ng + j)] + j #- grid - 1 # NOTICE
KmD[i,j] = mD[i : (m - ng + i), j : (n - ng + j)]
S = np.zeros_like(KmD[0,0])
Y = np.zeros_like(KmD[0,0])
# Inverse distance weighted
for i in range(ng):
for j in range(ng):
s = 1/np.sqrt(KmX[i,j] * KmX[i,j] + KmY[i,j] * KmY[i,j])
Y = Y + s * KmD[i,j]
S = S + s
S[S == 0] = 1
out = np.zeros((m,n))
out[grid + 1 : -grid, grid + 1 : -grid] = Y/S
return out