**ORB: An efficient alternative to SIFT or SURF**
We choose a large natural image to get image slices to simulate with¶
Yosemite_NP_M (1880 x 2880 x 3)
main_img = np.asarray(PIL.Image.open("Yosemite_NP_M.jpg"))
show_img(main_img)
Set up an Orb matcher¶
orb = cv2.ORB_create()
bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
FLANN_INDEX_KDTREE = 1
FLANN_INDEX_LSH = 6
draw_matches_params = dict(matchColor = (0,255,0),
singlePointColor = (255,0,0),
flags=cv2.DrawMatchesFlags_NOT_DRAW_SINGLE_POINTS)
# index_params = dict(algorithm = FLANN_INDEX_KDTREE, trees = 5)
index_params= dict(algorithm = FLANN_INDEX_LSH, table_number = 6, key_size = 12, multi_probe_level = 1)
search_params = dict(checks = 50)
flann = cv2.FlannBasedMatcher(index_params, search_params)
BF = True
FLANN = False
draw_params = dict(matchColor = (0,255,0), # draw matches in green color
singlePointColor = None,
# draw only inliers
flags = 2)
Run orb matching tracking and image transform across all the frame in the trajectory¶
traj = Trajectory(np.asarray(PIL.Image.open("Yosemite_NP_M.jpg")), 50)
start_img_copy = traj.get_first(scale=1.25)
start_img_copy = cv2.resize(start_img_copy, (600,400))
start_img = np.copy(start_img_copy)
# show_img(start_img,figsize=(10,6))
move_calc = np.zeros((traj.xyza.shape[0], 3))
move_true = np.zeros((traj.xyza.shape[0], 3))
num_kpts = np.zeros(traj.xyza.shape[0], dtype=np.int0)
start_img_gray = cv2.cvtColor(start_img, cv2.COLOR_BGR2GRAY)
kp2, des2 = orb.detectAndCompute(start_img_gray,None)
with FFMPEG_VideoWriter('orb_vid_stab.mp4', (900,300), 24.0) as video:
for i,f in enumerate(traj):
cur_img, move_true[i,:] = f
start_img = np.copy(start_img_copy)
# find the keypoints and descriptors with ORB
gray1 = cv2.cvtColor(cur_img, cv2.COLOR_BGR2GRAY)
# gray2 = cv2.cvtColor(start_img, cv2.COLOR_BGR2GRAY)
kp1, des1 = orb.detectAndCompute(gray1,None)
# kp2, des2 = orb.detectAndCompute(gray2,None)
if BF:
# matches = bf.knnMatch(des1,des2,k=2)
matches = bf.match(des1,des2)
good = sorted(matches, key = lambda x:x.distance)
num_kpts[i] = len(good)
elif FLANN:
matches = flann.knnMatch(des1,des2,k=2)
good = []
try:
for m,n in matches:
if m.distance < 0.7*n.distance:
good.append(m)
except:
print('Error',len(good))
MIN_MATCH_COUNT = 10
if len(good)>MIN_MATCH_COUNT:
src_pts = np.float32([ kp1[m.queryIdx].pt for m in good ]).reshape(-1,1,2)
dst_pts = np.float32([ kp2[m.trainIdx].pt for m in good ]).reshape(-1,1,2)
M, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC,5.0)
matchesMask = mask.ravel().tolist()
h,w,d = start_img.shape
pts = np.float32([ [0,0],[0,h-1],[w-1,h-1],[w-1,0] ]).reshape(-1,1,2)
dst = cv2.perspectiveTransform(pts,M)
start_img = cv2.polylines(start_img,[np.int32(dst)],True,(0,0,255),2, cv2.LINE_AA)
else:
print( "Not enough matches are found - {}/{}".format(len(good), MIN_MATCH_COUNT) )
matchesMask = None
img_kps = cv2.drawMatches(cur_img,kp1,start_img,kp2,good,None, matchesMask = matchesMask, **draw_params)
num_kpts[i] = len(good)
move_calc[i,:] = 0,0,0
cv2.imshow("keypoints", cv2.cvtColor(img_kps, cv2.COLOR_RGB2BGR))
cv2.waitKey(20)
img_kps = cv2.resize(img_kps, (900,300))
video.write_frame(img_kps)
print(':', end='', flush=True)
cv2.destroyAllWindows()
Show result video¶
mvp.ipython_display('orb_vid_stab.mp4', loop=True)
Show pandas table of accuracy¶
move_true[:,:2] = move_true[:,:2] - traj.startpos
df = pandas.DataFrame(np.column_stack((move_true, move_calc, num_kpts)))
df.columns = ['true_x','true_y', 'true_a','calc_y','calc_x', 'calc_a', 'num_kpts']
df.head(10)
plot accuracies¶
# out.release()
cum_mov_true = np.cumsum(move_true, axis=0)
cum_mov_calc = np.cumsum(move_calc, axis=0)
fig, axes = plt.subplots(1, 2, figsize=(10, 3))
axes[0].plot(cum_mov_true[:,0], cum_mov_true[:,1])
axes[0].set_title('Ground Truth trajectory')
axes[1].plot(cum_mov_calc[:,1], cum_mov_calc[:,0])
axes[1].set_title('Calculated Trajectory')
plt.tight_layout()
plt.show()
