第一次上传

Author: 曾锐

README.md

@@ -1,2 +1,36 @@
 # Hessian_sift
-Hessian-Affine detector, SIFT descriptor
+Hessian-Affine detector, SIFT descriptor  
+original from : [hesaff](https://github.com/perdoch/hesaff)
+
+
+## 编译特征提取接口
+Makefile(linux, opencv3.2)
+```
+all: *.cpp
+	g++ -O3 -Wall -o hesaff pyramid.cpp affine.cpp siftdesc.cpp helpers.cpp hesaff.cpp `pkg-config opencv4 --cflags --libs` -lrt
+
+```
+Makefile(mac, opencv4.1)
+```
+all: *.cpp
+	g++ -std=c++11 -O3 -Wall -o hesaff pyramid.cpp affine.cpp siftdesc.cpp helpers.cpp hesaff.cpp `pkg-config opencv4 --cflags --libs`
+```
+
+## python测试
+简单实验了原始sift和hesaff（图像宽最大限制500）
+```
+python demo_python.py
+```
+其中二进制文件(hesaff.sift)读取：[by insikk](https://github.com/perdoch/hesaff/blob/fcd04c9b7d2b7361ec676e1ded228768823eb0b3/util/Batch%20SIFT%20Extractor.ipynb)
+
+
+
+## Citation
+```
+Perdoch, M. and Chum, O. and Matas, J.: Efficient Representation of
+Local Geometry for Large Scale Object Retrieval. In proceedings of
+CVPR09. June 2009.
+
+TBD: A reference to technical report describing the details and some
+retrieval results will be placed here.
+```

demo_python.py

@@ -0,0 +1,122 @@
+import os
+import cv2
+import time
+import imutils
+import numpy as np
+sift = cv2.xfeatures2d.SIFT_create()
+
+
+def parse_sift_output(target_path):
+    """
+    Return:
+        kp: keypoint of hessian affine descriptor. location, orientation etc... OpenCV KeyPoint format. 
+        des: 128d uint8 np array
+    """
+    
+    # print(os.listdir("./sample"))
+    kp = []
+    des = []
+    with open(target_path, "r") as f:
+        lines = list(map(lambda x: x.strip(), f.readlines()))
+        num_descriptor = int(lines[1])
+        lines = lines[2:]
+        for i in range(num_descriptor):
+            # print(i, lines[i])
+            val = lines[i].split(" ")
+            x = float(val[0])
+            y = float(val[1])
+            a = float(val[2])
+            b = float(val[3])
+            c = float(val[4])
+            # TODO: generate ellipse shaped key point
+            # Refer: https://math.stackexchange.com/questions/1447730/drawing-ellipse-from-eigenvalue-eigenvector
+            # Refer: http://www.robots.ox.ac.uk/~vgg/research/affine/det_eval_files/display_features.m
+            # Refer: http://www.robots.ox.ac.uk/~vgg/research/affine/detectors.html
+            key_point = cv2.KeyPoint(x, y, 1)
+            sift_descriptor = np.array(list(map(lambda x: int(x), val[5:])), dtype=np.uint8)
+            kp.append(key_point)
+            des.append(sift_descriptor)
+        
+    
+    return kp, np.array(des)
+
+
+def resize(img):
+    h, w, _ = img.shape
+    if w > 500:
+        img = imutils.resize(img, width=500)
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+    return img
+
+def hessian_sift_extractor(path1, path2):
+    os.system('./hesaff_c++/hesaff {}'.format(path1))
+    os.system('./hesaff_c++/hesaff {}'.format(path2))
+
+    kp1, des1 = parse_sift_output(path1 + '.hesaff.sift')
+    kp2, des2 = parse_sift_output(path2 + '.hesaff.sift')
+    des1 = des1.astype('float32')
+    des2 = des2.astype('float32')
+    return kp1, des1, kp2, des2
+
+def sift_extractor(img1, img2):
+    kp1, des1 = sift.detectAndCompute(img1, None)
+    kp2, des2 = sift.detectAndCompute(img2, None)
+    return kp1, des1, kp2, des2
+
+
+def show_match(path1, path2):
+
+    img1 = cv2.imread(path1, 1)
+    img2 = cv2.imread(path2, 1)
+    img1_gray  = resize(img1.copy())
+    img2_gray  = resize(img2.copy())
+    
+    kp1, des1, kp2, des2 = hessian_sift_extractor(path1, path2)
+    # kp1, des1, kp2, des2 = sift_extractor(img1_gray, img2_gray)
+
+    #### FLANN
+    FLANN_INDEX_KDTREE = 0
+    index_params = dict(algorithm = FLANN_INDEX_KDTREE, trees = 5)
+    search_params = dict(checks = 50) #指定递归次数
+
+    begin = time.time()
+    flann = cv2.FlannBasedMatcher(index_params, search_params)
+    matches = flann.knnMatch(des1, des2, k=2)
+
+    # Need to draw only good matches, so create a mask
+    matchesMask = [[0, 0] for i in range(len(matches))]
+    
+
+    del_same = []
+    mat = 0
+    for i, (m, n) in enumerate(matches):
+        # ratio test as per Lowe's paper
+        if m.distance < 0.6 * n.distance:
+            mat += 1
+            matchesMask[i] = [1, 0]
+            del_same.append(m.trainIdx)
+    count = len(set(del_same))
+    print('匹配耗时：{:.5f}, 匹配到的点个数：{}'.format(time.time() - begin, count))
+
+
+
+
+    draw_params = dict(matchColor = (0, 255, 0),
+                       singlePointColor = (255, 0, 0),
+                       matchesMask = matchesMask,
+                       flags = 0)
+    img3 = cv2.drawMatchesKnn(img1_gray, kp1, img2_gray, kp2, matches, None, **draw_params)
+    # img3 = cv2.drawKeypoints(img1, kp1, img1,flags=cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS) #画出特征点，并显示为红色圆圈
+
+    cv2.namedWindow('img',cv2.WINDOW_AUTOSIZE)
+    cv2.imshow('img',img3)
+    cv2.waitKey(0)
+    cv2.destroyAllWindows()
+    # cv2.imwrite('test.jpg', img3)
+
+
+if __name__ == '__main__':
+
+    path1 = './test_pics/1.jpg'
+    path2 = './test_pics/2.jpg'
+    show_match(path1, path2)

hesaff_c++

@@ -0,0 +1 @@
+Subproject commit 0d939a424372ed7dd23ecc0b3590f37122695cbc