2021 lab

Lab_1/Readme.md

@@ -22,6 +22,13 @@ You can only use the python packages below:
  - sklearn
  - math
 
+## Evaluation
+For your output image, you could use eval.py to check your answer.
+```
+$ python3 eval.py --gt GT/1-1/fig_grad_x.jpg --out output/1-1/fig_grad_y.jpg
+```
+It will show that you "PASS!" or "FAIL." this question.
+
 ## 1. Image Filtering
 1.	Load "**fig.jpg**" with **gray** scale.
 2.	Compute and save the results of gradient generated by horizontal Sobel filter and vertical Sobel filter as **fig_grad_x.jpg** and **fig_grad_y.jpg** (with cv2.filter2D).
@@ -32,11 +39,11 @@ The dataset you need is under `Face_dataset/` directory, which contains 56×46 p
 
 1. Perform PCA on the **training set**. Save the **mean face** and the first **three** eigenfaces.
 	- save to [mean.png], [1.png], [2.png], [3.png]
-	- Be careful the values in eigenvectors after PCA !  (*cv2.PCACompute()*)
+	- Be careful the values in eigenvectors after PCA !  (*sklearn.decomposition.PCA*, you don't need to specify n_components value)
 	- Bonus: you can implement your own PCA function !!
 	- hint: 　255 * (shifted **x**/ max(shifted **x**))
 
-2. Take **7_2.png**, and project it onto the above PCA eigenspace. Reconstruct this image using the first n = 3, 100 eigenfaces. For each n, compute the mean square error (MSE) between the reconstructed face image and **7_2.png**. Please save these reconstructed images, and record the corresponding MSE values.
+2. Take **7_2.png**, and project it onto the above PCA eigenspace. Reconstruct this image using the first n = 3, 100 eigenfaces. For each n, compute the mean square error (MSE) between the reconstructed face image and **7_2.png**. Please save these reconstructed images , [7_2_n3.png] and [7_2_n100.png], and record the corresponding MSE values.
 
 3. Apply the **k-nearest neighbors classifier** to recognize **test set** images. For simplicity, try k={1,3} and n = {3,100}.  Report the recognition rate on the test set with different (k,n) pairs.
 	- hint: *sklearn.neighbors.KNeighborsClassifier()*

Lab_1/eval.py

@@ -0,0 +1,19 @@
+import os
+import numpy as np
+import argparse
+import cv2
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--gt', help='ground truth path')
+    parser.add_argument('--out', help='your output file path')
+
+    args = parser.parse_args()
+    try:
+        if (cv2.imread(os.path.join(args.gt)) - cv2.imread(os.path.join(args.out))).sum() == 0.0: print("PASS!")
+        else: print("FAIL!")
+    except:
+        print("FAIL!")
+
+if __name__ == '__main__':
+    main()

Lab_2/main.py

@@ -1,28 +1,86 @@
-import numpy as np
-import numpy.linalg as LA
-import cv2
-
-
-# u, v are N-by-2 matrices, representing N corresponding points for v = T(u)
-# this function should return a 3-by-3 homography matrix
-def solve_homography(u, v):
-	# TODO
-	return H
-
-def main():
-
-    # Input Initialization 
-	corners = np.array([[190, 106], [256, 142], [193, 244], [258, 285]]) # Points [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
-	img_path = './screen.jpg'
-
-    # Output initialization (create an output array to save the result)
-	new_h, new_w = 300, 300
-
-    # TODO Solving homography matrix for backward warping (hints: instead of solve v = Hu, solve u = (H^-1)v)
-
-    # TODO Backward Warping 
-
-
-
-if __name__ == '__main__':
-	main()
+import numpy as np
+import numpy.linalg as LA
+import cv2
+
+
+# u, v are N-by-2 matrices, representing N corresponding points for v = T(u)
+# this function should return a 3-by-3 homography matrix
+def solve_homography(u, v):
+    """
+    This function should return a 3-by-3 homography matrix,
+    u, v are N-by-2 matrices, representing N corresponding points for v = T(u)
+    :param u: N-by-2 source pixel location matrices
+    :param v: N-by-2 destination pixel location matrices
+    :return:
+    """
+    # TODO: 1.forming A
+
+    # TODO: 2.solve H with A
+
+    return H
+
+def warping(src, dst, H, ymin, ymax, xmin, xmax):
+    """
+    Perform forward/backward warpping without for loops. i.e.
+    for all pixels in src(xmin~xmax, ymin~ymax),  warp to destination
+          (xmin=0,ymin=0)  source                       destination
+                         |--------|              |------------------------|
+                         |        |              |                        |
+                         |        |     warp     |                        |
+    forward warp         |        |  --------->  |                        |
+                         |        |              |                        |
+                         |--------|              |------------------------|
+                                 (xmax=w,ymax=h)
+
+    for all pixels in dst(xmin~xmax, ymin~ymax),  sample from source
+                            source                       destination
+                         |--------|              |------------------------|
+                         |        |              | (xmin,ymin)            |
+                         |        |     warp     |           |--|         |
+    backward warp        |        |  <---------  |           |__|         |
+                         |        |              |             (xmax,ymax)|
+                         |--------|              |------------------------|
+
+    :param src: source image
+    :param dst: destination output image
+    :param H:
+    :param ymin: lower vertical bound of the destination(source, if forward warp) pixel coordinate
+    :param ymax: upper vertical bound of the destination(source, if forward warp) pixel coordinate
+    :param xmin: lower horizontal bound of the destination(source, if forward warp) pixel coordinate
+    :param xmax: upper horizontal bound of the destination(source, if forward warp) pixel coordinate
+    :param direction: indicates backward warping or forward warping
+    :return: destination output image
+    """
+
+    # TODO: 1.meshgrid the (x,y) coordinate pairs
+
+    # TODO: 2.reshape the destination pixels as N x 3 homogeneous coordinate
+
+    # TODO: 3.apply H_inv to the destination pixels and retrieve (u,v) pixels, then reshape to (ymax-ymin),(xmax-xmin)
+
+    # TODO: 4.calculate the mask of the transformed coordinate (should not exceed the boundaries of source image)
+
+    # TODO: 5.sample the source image with the masked and reshaped transformed coordinates
+
+    # TODO: 6. assign to destination image with proper masking
+
+
+    return dst
+
+def main():
+
+    # Input Initialization 
+	corners = np.array([[190, 106], [256, 142], [193, 244], [258, 285]]) # Points [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
+	img_path = './screen.jpg'
+
+    # Output initialization (create an output array to save the result)
+	new_h, new_w = 300, 300
+
+    # TODO Solving homography matrix for backward warping (hints: instead of solve v = Hu, solve u = (H^-1)v)
+
+    # TODO Backward Warping 
+
+
+
+if __name__ == '__main__':
+	main()