Add multi-channel support with hog

CSRT/CSRT.py

@@ -1,7 +1,4 @@
 import cv2
-from skimage import feature
-from skimage.feature import hog
-from skimage.transform import resize
 import HOG_features as hogfeat
 import numpy as np
 
@@ -34,13 +31,11 @@ def cap_func(element):
 class CSRT():
 
     def __init__(self, frame, roi, num_features, debug = False):
-        self.debug = None
+        self.debug = debug
 
         self.frame = frame
         self.roi = roi
-        self.sigma = 100
-        self.mu = 5
-        self.beta, self.λ, self.n = 3, 0.01, 0.02
+        self.beta, self.λ, self.n, self.sigma, self.mu = 3, 0.01, 0.02, 100, 5
         self.g = get_gaussian_map(self.roi, self.sigma)
 
         if self.debug:
@@ -51,12 +46,12 @@ def __init__(self, frame, roi, num_features, debug = False):
         self.p = [] # This variable will store position of object in each frame.
         self.channel_weights = [0] * num_features # Will store channel weights for all channels.
         self.G_cap = [0] * num_features # Will store individual G_cap/g_tilda values for channels.
-        self.G_res = 0 # Will store resultant G_cap after using channel_reliability.
+        # self.G_res = 0 # Will store resultant G_cap after using channel_reliability.
 
 
     def set_roiImage(self):
-        x,y,w,h = self.roi
-        self.roi_img = self.frame[y:y+h, x:x+w]
+        self.x, self.y, self.width, self.height = self.roi
+        self.roi_img = self.frame[self.y : self.y + self.height, self.x : self.x + self.width]
         self.h = np.zeros((self.roi_img.shape[0], self.roi_img.shape[1]))
         self.I = np.zeros_like(self.h)
 
@@ -136,9 +131,9 @@ def update_H(self):
 
                 self.mu_i *= self.beta
             self.h_cap[channel_index] = h_cap
-        print("h_cap for feature channels ", self.h_cap)
-        print("length h_cap", len(self.h_cap))
-        print("Lengths of h_cap", len(h_cap[0]), len(h_cap[1]))
+        # print("h_cap for feature channels ", self.h_cap)
+        # print("length h_cap", len(self.h_cap))
+        # print("Lengths of h_cap", len(h_cap[0]), len(h_cap[1]))
 
 
     def calculate_g_cap_and_channel_weights(self):
@@ -167,19 +162,36 @@ def calculate_g_cap_and_channel_weights(self):
             sum += self.channel_weights[channel_index]
         for channel_index in range(len(self.features)):
             self.channel_weights[channel_index] /= sum
-        print("channel_weights", self.channel_weights)
-        print("G_cap", self.G_cap)
+        # print("channel_weights", self.channel_weights)
+        # print("G_cap", self.G_cap)
 
         if self.debug:
             print("roi", self.roi)
             print('G shape', self.features[0].shape)
 
-def calculate_final_g_cap(self):
-    for channel_index in range(len(self.features)):
-        self.G_res += self.G_cap[channel_index] * self.channel_weights[channel_index]
-    return self.G_res
-
-def apply_csrt(self):
-    self.p = self.get_new_roi()
-    # print(self.p)
-    self.channel_reliability()
+    def calculate_final_g_cap(self):
+        for channel_index in range(len(self.features)):
+            self.G_res += self.G_cap[channel_index] * self.channel_weights[channel_index]
+        return self.G_res
+
+    def draw_bbox(self):
+        self.G_res = self.G_cap[0] * self.channel_weights[0]
+        for channel_index in range(1, len(self.features)):
+            self.G_res += self.G_cap[channel_index] * self.channel_weights[channel_index]
+        self.G_res = (self.G_res - self.G_res.min()) / (self.G_res.max() - self.G_res.min())
+        max_value = np.max(self.G_res)
+        max_pos = np.where(self.G_res == max_value)
+        dy = int(np.mean(max_pos[0]) - self.G_res.shape[0] / 2)
+        dx = int(np.mean(max_pos[1]) - self.G_res.shape[1] / 2)
+        self.p = [self.x + dx, self.y + dy, self.x + dx + self.width, self.y + dy + self.height]
+        print(self.p[0])
+        print(self.p[1])
+        print(self.p[2])
+        print(self.p[3])
+        cv2.rectangle(self.frame, (self.p[0], self.p[1]), (self.p[2], self.p[3]), (0, 255, 0), 2)
+        cv2.imshow("frame", self.frame)
+        cv2.waitKey(0)
+        cv2.destroyAllWindows()
+
+    def apply_csrt(self):
+        pass

CSRT/HOG_features.py

@@ -1,9 +1,15 @@
 import cv2
 from skimage.feature import hog
-# from skimage.transform import resize
+import numpy as np
 
 def get_hog_features(image, des_orientations, des_pixels_per_cell):
     fd, hog_image = hog(image, orientations = des_orientations,
         pixels_per_cell = (des_pixels_per_cell, des_pixels_per_cell), cells_per_block=(2, 2),
         visualize=True,multichannel=True)
+    # np.resize(fd, hog_image.shape)
+    # print("fd shape", fd.shape)
+    # print("hog shape", hog_image.shape)
+    # cv2.imshow("Hog image", hog_image)
+    # cv2.waitKey(0)
+    # cv2.destroyAllWindows()
     return hog_image

CSRT/test.py

@@ -1,5 +1,4 @@
 import cv2
-from skimage.transform import resize
 from CSRT import *
 
 
@@ -10,12 +9,25 @@
 # print(type(img))
 # print(img.shape)
 
-tracker = CSRT(img, roi, 2)
+tracker = CSRT(img, roi, 20)
 tracker.set_roiImage()
 tracker.get_spatial_reliability_map()
 tracker.generate_features(8, 2)
 tracker.update_H()
 tracker.calculate_g_cap_and_channel_weights()
+tracker.draw_bbox()
+
+# video = cv2.VideoCapture("/home/prathamesh/OpenCV-Trackers/assets/chaplin.mp4")
+
+# tracker = CSRT(5)
+# _, frame1 = video.read()
+# roi = cv2.selectROI('Select Roi', frame1)
+# tracker.next_frame(frame1, roi)
+
+# while(video.isOpened()):
+#     pass
+
+
 # tracker.apply_csrt()
 
 # x, y, w, h = tracker.get_new_roi()