Image stitching problem using Python and OpenCV

Question

I got output like below after stitching result of 24 stitched images to next 25th image. Before that stitching was good.

Is anyone aware of why/when output of stitching comes like this? What are the possibilities of output coming like that? What may be the reason of that?

Stitching code is following standard stitching steps like finding keypoints, descriptors then matching points, calculating homography and then warping of images. But I am not understanding why that output is coming.

Core part of stitching is like below:

detector = cv2.SIFT_create(400)
# find the keypoints and descriptors with SIFT
gray1 = cv2.cvtColor(image1,cv2.COLOR_BGR2GRAY)
ret1, mask1 = cv2.threshold(gray1,1,255,cv2.THRESH_BINARY)
kp1, descriptors1 = detector.detectAndCompute(gray1,mask1)

gray2 = cv2.cvtColor(image2,cv2.COLOR_BGR2GRAY)
ret2, mask2 = cv2.threshold(gray2,1,255,cv2.THRESH_BINARY)
kp2, descriptors2 = detector.detectAndCompute(gray2,mask2)

keypoints1Im = cv2.drawKeypoints(image1, kp1, outImage = cv2.DRAW_MATCHES_FLAGS_DEFAULT, color=(0,0,255))
keypoints2Im = cv2.drawKeypoints(image2, kp2, outImage = cv2.DRAW_MATCHES_FLAGS_DEFAULT, color=(0,0,255))

# BFMatcher with default params
matcher = cv2.BFMatcher()
matches = matcher.knnMatch(descriptors2,descriptors1, k=2)

# Apply ratio test
good = []
for m, n in matches:
    if m.distance < 0.75 * n.distance:
        good.append(m)

print (str(len(good)) + " Matches were Found")

if len(good) <= 10:
    return image1

matches = copy.copy(good)

matchDrawing = util.drawMatches(gray2,kp2,gray1,kp1,matches)

#Aligning the images
src_pts = np.float32([ kp2[m.queryIdx].pt for m in matches ]).reshape(-1,1,2)
dst_pts = np.float32([ kp1[m.trainIdx].pt for m in matches ]).reshape(-1,1,2)


H = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC,5.0)[0]

h1,w1 = image1.shape[:2]
h2,w2 = image2.shape[:2]
pts1 = np.float32([[0,0],[0,h1],[w1,h1],[w1,0]]).reshape(-1,1,2)
pts2 = np.float32([[0,0],[0,h2],[w2,h2],[w2,0]]).reshape(-1,1,2)
pts2_ = cv2.perspectiveTransform(pts2, H)
pts = np.concatenate((pts1, pts2_), axis=0)
# print("pts:", pts)
[xmin, ymin] = np.int32(pts.min(axis=0).ravel() - 0.5)
[xmax, ymax] = np.int32(pts.max(axis=0).ravel() + 0.5)
t = [-xmin,-ymin]
Ht = np.array([[1,0,t[0]],[0,1,t[1]],[0,0,1]]) # translate

result = cv2.warpPerspective(image2, Ht.dot(H), (xmax-xmin, ymax-ymin))

resizedB = np.zeros((result.shape[0], result.shape[1], 3), np.uint8)

resizedB[t[1]:t[1]+h1,t[0]:w1+t[0]] = image1
# Now create a mask of logo and create its inverse mask also
img2gray = cv2.cvtColor(result,cv2.COLOR_BGR2GRAY)
ret, mask = cv2.threshold(img2gray, 0, 255, cv2.THRESH_BINARY)

kernel = np.ones((5,5),np.uint8)
k1 = (kernel == 1).astype('uint8')
mask = cv2.erode(mask, k1, borderType=cv2.BORDER_CONSTANT)

mask_inv = cv2.bitwise_not(mask)

difference = cv2.bitwise_or(resizedB, resizedB, mask=mask_inv)

result2 = cv2.bitwise_and(result, result, mask=mask)

result = cv2.add(result2, difference)

Edit:

This image shows match drawing while stitching 25 to result until 24 images:

And before that match drawing:

I have total 97 images to stitch. If I stitch 24 and 25 image separately they stitches properly. If I start stitching from 23rd image onwards then also stitching is good but it gives me problem when I stitches starting from 1st image. I am not able to understand the problem.

Result after stitching 23rd image:

Result after stitching 24th image:

Result after stitching 25th image is as above which went wrong.

Strange Observation: If I stitch 23,24,25 images seperately with same code it gets stitches. If I stitch images after 23 till 97 , it gets stitches. But somehow if I stitch images from 1st, it breaks while stitching 25th image. I am not understanding why this happens.

I have tried different combination like different keypoint detection, extraction methods, matching methods, different homography calculations, different warping code but those combinations didn't work. Something is missing or wrong in the steps combination code. I am not able to figure it out.

Sorry for this long question. As I am completely new to this I am not able to explain and get the things properly. Thanks for your help and guidance.

Stitched result of 23,24,25 images separately with SAME code:

With different code (gives black lines in between stitching), if I stitched 97 images then 25th goes up in stitching and stitches as shown below (right corner point):

Answer 1

Firstly, I was not able to recreate your problem and solve it as the images were too big for my system to process. However, I had faced the same problem in my Panorama Stitching project, so I am sharing the reason behind it and my approach to solving my problem. Hope this helps you too.

Here's what my problem looked like when I stitched 4 images together just like you did.

As you can see, the 4th image was getting distorted a lot which must not happen. The same thing happened with you but on a greater level.

Now, here's the output when I stitched 8 images after some image pre-processing.

After some pre-processing on the input images, I was able to stitch 8 images together perfectly without any distortion.

To understand the exact reason behind this kind of distortion, watch this video by Joseph Redmon between 50:26 - 1:07:23.

As suggested in the video, we'll first have to project the images onto a cylinder and then unroll them and then stitch these unrolled images together.

Below is the initial input image(left) and the image after projection and unrolling onto a cylinder(right).

For your problem, as you are using satellite images, I guess projection onto a sphere would work better than the cylinder however you'll have to give it a try.

Sharing below my code for projecting the image onto a cylinder and unrolling it for reference. The mathematics used behind it is the same as given in the video.

def Convert_xy(x, y):
    global center, f

    xt = ( f * np.tan( (x - center[0]) / f ) ) + center[0]
    yt = ( (y - center[1]) / np.cos( (x - center[0]) / f ) ) + center[1]
    
    return xt, yt


def ProjectOntoCylinder(InitialImage):
    global w, h, center, f
    h, w = InitialImage.shape[:2]
    center = [w // 2, h // 2]
    f = 1100       # 1100 field; 1000 Sun; 1500 Rainier; 1050 Helens
    
    # Creating a blank transformed image
    TransformedImage = np.zeros(InitialImage.shape, dtype=np.uint8)
    
    # Storing all coordinates of the transformed image in 2 arrays (x and y coordinates)
    AllCoordinates_of_ti =  np.array([np.array([i, j]) for i in range(w) for j in range(h)])
    ti_x = AllCoordinates_of_ti[:, 0]
    ti_y = AllCoordinates_of_ti[:, 1]
    
    # Finding corresponding coordinates of the transformed image in the initial image
    ii_x, ii_y = Convert_xy(ti_x, ti_y)

    # Rounding off the coordinate values to get exact pixel values (top-left corner)
    ii_tl_x = ii_x.astype(int)
    ii_tl_y = ii_y.astype(int)

    # Finding transformed image points whose corresponding 
    # initial image points lies inside the initial image
    GoodIndices = (ii_tl_x >= 0) * (ii_tl_x <= (w-2)) * \
                  (ii_tl_y >= 0) * (ii_tl_y <= (h-2))

    # Removing all the outside points from everywhere
    ti_x = ti_x[GoodIndices]
    ti_y = ti_y[GoodIndices]
    
    ii_x = ii_x[GoodIndices]
    ii_y = ii_y[GoodIndices]

    ii_tl_x = ii_tl_x[GoodIndices]
    ii_tl_y = ii_tl_y[GoodIndices]

    # Bilinear interpolation
    dx = ii_x - ii_tl_x
    dy = ii_y - ii_tl_y

    weight_tl = (1.0 - dx) * (1.0 - dy)
    weight_tr = (dx)       * (1.0 - dy)
    weight_bl = (1.0 - dx) * (dy)
    weight_br = (dx)       * (dy)
    
    TransformedImage[ti_y, ti_x, :] = ( weight_tl[:, None] * InitialImage[ii_tl_y,     ii_tl_x,     :] ) + \
                                      ( weight_tr[:, None] * InitialImage[ii_tl_y,     ii_tl_x + 1, :] ) + \
                                      ( weight_bl[:, None] * InitialImage[ii_tl_y + 1, ii_tl_x,     :] ) + \
                                      ( weight_br[:, None] * InitialImage[ii_tl_y + 1, ii_tl_x + 1, :] )


    # Getting x coorinate to remove black region from right and left in the transformed image
    min_x = min(ti_x)

    # Cropping out the black region from both sides (using symmetricity)
    TransformedImage = TransformedImage[:, min_x : -min_x, :]

    return TransformedImage, ti_x-min_x, ti_y

You just have to call the function ProjectOntoCylinder and pass it an image to get the resultant image and the coordinates of white pixels in the mask image. Use the code below to call this function and get the mask image.

# Applying Cylindrical projection on Image
Image_Cyl, mask_x, mask_y = ProjectOntoCylinder(Image)

# Getting Image Mask
Image_Mask = np.zeros(Image_Cyl.shape, dtype=np.uint8)
Image_Mask[mask_y, mask_x, :] = 255

Here are links to my project and its detailed documentation for reference:

Part 1: Source Code, Documentation

Part 2: Source Code, Documentation