def calculate_point(kps_left, sco_left, des_left, kps_right, sco_right, des_right):
FLANN_INDEX_KDTREE = 1
index_params = dict(algorithm=FLANN_INDEX_KDTREE, trees=5)
search_params = dict(checks=50)
flann = cv2.FlannBasedMatcher(index_params, search_params)
matches = flann.knnMatch(des_left, des_right, k=2)
# matcher = cv2.DescriptorMatcher_create(cv2.DescriptorMatcher_FLANNBASED)
# matches = matcher.knnMatch(des_left, des_right, 2)
goodMatch = []
locations_1_to_use = []
locations_2_to_use = []
# 匹配对筛选
min_dist = 1000
max_dist = 0
disdif_avg = 0
# 统计平均距离差
for m, n in matches:
disdif_avg += n.distance - m.distance
disdif_avg = disdif_avg / len(matches)
# print('disdif_avg:', disdif_avg)
for m, n in matches:
#自适应阈值
if n.distance > m.distance + 1*disdif_avg:
# if m.distance < 0.9 * n.distance:
goodMatch.append(m)
p2 = cv2.KeyPoint(kps_right[m.trainIdx][0], kps_right[m.trainIdx][1], 1)
p1 = cv2.KeyPoint(kps_left[m.queryIdx][0], kps_left[m.queryIdx][1], 1)
locations_1_to_use.append([p1.pt[0], p1.pt[1]])
locations_2_to_use.append([p2.pt[0], p2.pt[1]])
#goodMatch = sorted(goodMatch, key=lambda x: x.distance)
# print('match num is %d' % len(goodMatch))
locations_1_to_use = np.array(locations_1_to_use)
locations_2_to_use = np.array(locations_2_to_use)
# Perform geometric verification using RANSAC.
_, inliers = measure.ransac((locations_1_to_use, locations_2_to_use),
transform.AffineTransform,
min_samples=3,
residual_threshold=_RESIDUAL_THRESHOLD,
max_trials=1000)
print('Found %d inliers' % sum(inliers))
inlier_idxs = np.nonzero(inliers)[0]
#最终匹配结果
matches = np.column_stack((inlier_idxs, inlier_idxs))
# print('whole time is %6.3f' % (time.perf_counter() - start0))
# Visualize correspondences, and save to file.
#1 绘制匹配连线
plt.rcParams['savefig.dpi'] = 100 #图片像素
plt.rcParams['figure.dpi'] = 100 #分辨率
plt.rcParams['figure.figsize'] = (4.0, 3.0) # 设置figure_size尺寸
_, ax = plt.subplots()
plotmatch.plot_matches(
ax,
image1,
image2,
locations_1_to_use,
locations_2_to_use,
np.column_stack((inlier_idxs, inlier_idxs)),
plot_matche_points = False,
matchline = True,
matchlinewidth = 0.5)
ax.axis('off')
ax.set_title('')
plt.show()
# print('inlier_idxs:', len(inlier_idxs))
res = locations_2_to_use[inlier_idxs] - locations_1_to_use[inlier_idxs]
return locations_1_to_use[inlier_idxs], locations_2_to_use[inlier_idxs]
def __init__(self, sift_kp_size=16, concat=True, kps_idx=numpy.arange(4, 31, 8)):
self.concat = concat
self.sift = cv2.SIFT()
coordinates = list(itertools.product(kps_idx, kps_idx))
self.kps = [cv2.KeyPoint(i, j, sift_kp_size) for (i, j) in coordinates]
def to_cv2_kp(kp):
# assert kp = [<row>, <col>, <ori>, <octave_ind>, <layer_ind>]
ratio = get_size_ratio_by_octave(kp[3])
scale = get_scale_by_ind(kp[3], kp[4])
return cv2.KeyPoint(kp[1] / ratio, kp[0] / ratio, 10,
kp[2] / PI * 180)
def load_graffiti_help(graffiti_dir, N_imgs, N_data, thresh):
# Load graffiti dataset images
n = N_data
N_points = 10 * n
data_files = [
os.path.join(graffiti_dir, 'img{}.png'.format(i + 1))
for i in range(N_imgs)
]
imgs = [imageio.imread(fname)[:, :, :3] for fname in data_files]
# Build Homographies
H = [np.eye(3)]
for i in range(1, N_imgs):
h = []
with open(os.path.join(graffiti_dir, 'H1to{}p'.format(i + 1)),
'r') as f:
for l in f:
h.append([float(x) for x in l.split()])
H.append(np.array(h))
# Compute sift descriptors for first image
sift = cv2.xfeatures2d.SIFT_create(N_points)
keypoints0, descs0 = sift.detectAndCompute(imgs[0], None)
# Select features with valid x,y coordinates not too close to others
descs0 = [d / np.linalg.norm(d) for d in descs0[:N_points]]
xy0 = np.stack([[k.pt[0], k.pt[1], 1] for k in keypoints0], axis=1)
ids1 = distint_locs(xy0, thresh)
xy0 = np.stack([xy0[:, ii] for ii in ids1], axis=1)
sel = xy0[0, :] < np.inf # Should all be true
for j in range(1, N_imgs):
xyj = h_apply(H[j], xy0)
and_lists = [
xyj[0, :] > 0,
xyj[1, :] > 0,
xyj[0, :] < imgs[j].shape[1],
xyj[1, :] < imgs[j].shape[0],
]
sel = list(functools.reduce(np.logical_and, and_lists, sel))
# Sort features by their selectiveness
descs0 = np.stack([descs0[ids1[i]] for i in range(len(ids1)) if sel[i]])
keypoints0 = [keypoints0[ids1[i]] for i in range(len(ids1)) if sel[i]]
Sim = np.dot(descs0, descs0.T)
SimSortList = [np.sort(s)[::-1] for s in Sim]
vec = (np.arange(len(descs0))[::-1])**(2)
best_idxs = sorted(np.arange(len(descs0)),
key=lambda x: SimSortList[x] @ vec)
# Select the features and pull out their features
xy_ = xy0[:, sel]
xy = [xy_[:, best_idxs[:n]]]
descs = [descs0[best_idxs[:n], :]]
keypoints = [[keypoints0[i] for i in best_idxs[:n]]]
# Compute the features for their location in the other images
for i in range(1, N_imgs):
xy.append(h_apply(H[i], xy[0]))
kpts = []
for j in range(xy[0].shape[1]):
x0, y0 = xy[0][0, j], xy[0][1, j]
size_new, angle_new = get_size_angle(x0, y0, H[i], keypoints0[j])
octv = size_to_octave_desc(size_new)
kpts.append(
cv2.KeyPoint(x=xy[i][0, j],
y=xy[i][1, j],
_size=size_new,
_angle=angle_new,
_response=keypoints0[j].response,
_octave=octv,
_class_id=keypoints0[j].class_id))
keypoints.append(kpts)
# Compute the full descriptors
# These include the feature descriptor, x, y, log scale, and orientation
FullDescs = []
for kpts, img in zip(keypoints, imgs):
h, w, c = imgs[i].shape
_, descs = sift.compute(img, kpts)
d_ = [(d / np.linalg.norm(d)) for d in descs]
dkp_ = [
np.array(
list(d_[i]) + [(k.pt[0] - w / 2) / w, (k.pt[1] - h / 2) / h,
np.log(k.size),
np.deg2rad(k.angle)])
for i, k in enumerate(kpts)
]
FullDescs.append(dkp_)
return np.stack(FullDescs)
x_0, y_0, _ = tuple(h_apply(H[0, i], (x0 + s * dx_, y0 + s * dy_, 1)))
x_1, y_1, _ = tuple(h_apply(H[0, i], (x0 - s * dy_, y0 + s * dx_, 1)))
x_2, y_2, _ = tuple(h_apply(H[0, i], (x0 + s * dy_, y0 - s * dx_, 1)))
x_3, y_3, _ = tuple(h_apply(H[0, i], (x0 - s * dx_, y0 - s * dy_, 1)))
s_new = np.mean([
np.linalg.norm((x_0 - x_3, y_0 - y_3)) / 2,
np.linalg.norm(((x_1 - x_2, y_1 - y_2))) / 2
])
octv = size_to_octave_desc(s_new)
angle_new = np.arctan2(-y_0 + y_3, x_0 - x_3)
angle_new = np.rad2deg(angle_new + 2 * np.pi * (angle_new < 0))
kpts.append(
cv2.KeyPoint(x=x,
y=y,
_size=s_new,
_angle=angle_new,
_response=keypoints0[j].response,
_octave=octv,
_class_id=keypoints0[j].class_id))
keypoints.append(kpts)
descriptors = []
Descs = []
FullDesc = []
for kpts, img in zip(keypoints, imgs):
h, w, c = imgs[i].shape
_, descs = sift.compute(img, kpts)
d_ = [(d / np.linalg.norm(d)) for d in descs]
descriptors.append(d_)
dkp_ = [
np.array(
def main(config):
# Build Networks
tf.reset_default_graph()
photo_ph = tf.placeholder(tf.float32, [1, None, None, 1]) # input grayscale image, normalized by 0~1
is_training = tf.constant(False) # Always False in testing
ops = build_networks(config, photo_ph, is_training)
tfconfig = tf.ConfigProto()
tfconfig.gpu_options.allow_growth = True
sess = tf.Session(config=tfconfig)
sess.run(tf.global_variables_initializer())
# load model
saver = tf.train.Saver()
print('Load trained models...')
if os.path.isdir(config.model):
checkpoint = tf.train.latest_checkpoint(config.model)
model_dir = config.model
else:
checkpoint = config.model
model_dir = os.path.dirname(config.model)
if checkpoint is not None:
print('Checkpoint', os.path.basename(checkpoint))
print("[{}] Resuming...".format(time.asctime()))
saver.restore(sess, checkpoint)
else:
raise ValueError('Cannot load model from {}'.format(model_dir))
print('Done.')
##########################################################################
new_size = (config.w, config.h)
# setup output dir
res_dir = os.path.join('res/lfnet/', config.trials)
if not os.path.exists(res_dir):
os.makedirs(res_dir)
# write human readable logs
f = open(os.path.join(res_dir, 'log.txt'), 'w')
f.write('lfnet\n')
f.write('data: %s\n'%cst.DATA)
f.write('thresh_overlap: %d\n'%cst.THRESH_OVERLAP)
f.write('thresh_desc: %d\n'%cst.THRESH_DESC)
# feature matcher handler for visualization
FLANN_INDEX_KDTREE = 0
index_params = dict(algorithm = FLANN_INDEX_KDTREE, trees = 5)
search_params = dict(checks=50)
matcher = cv2.FlannBasedMatcher(index_params,search_params)
norm = 'L2'
global_start_time = time.time()
for scene_name in cst.SCENE_LIST:
print('************ %s ************'%scene_name)
f.write('************ %s ************\n'%scene_name)
scene_dir = '%s/%s/'%(cst.DATA_DIR, scene_name)
pose_fn = '%s/images.txt'%(scene_dir)
img_list = [ l for l in sorted(os.listdir(scene_dir)) if l[-3:]=='png']
img_num = len(img_list)
# intrinsic params (same for all img)
camera_fn = os.path.join(scene_dir, '0000.png.camera')
#print('camera_fn: %s'%camera_fn)
K, T = bench_tools.load_camera(camera_fn)
fx = K[0,0]
fy = K[1,1]
cx = K[0,2]
cy = K[1,2]
K_inv = np.linalg.inv(K)
img0_root_fn = '%04d'%0
img0_fn = os.path.join(scene_dir, '%s.png'%img0_root_fn)
img0 = cv2.imread(img0_fn)
old_size = (img0.shape[1], img0.shape[0])
img0 = cv2.resize(img0, new_size, interpolation=cv2.INTER_AREA)
sx2big, sy2big = 1.0*old_size[0]/new_size[0], 1.0*old_size[1]/new_size[1]
sx2small, sy2small = 1.0*new_size[0]/old_size[0], 1.0*new_size[1]/old_size[1]
print('sx2big: %.3f - sy2big: %.3f'%(sx2big, sy2big))
print('sx2small: %.3f - sy2small: %.3f'%(sx2small, sy2small))
# get colmap depth map
depth0_fn = '%s/%s.png.photometric.bin'%(scene_dir, img0_root_fn)
depth0_map = bench_tools.read_array(depth0_fn)
depth0_map = cv2.resize(depth0_map, old_size, interpolation=cv2.INTER_CUBIC)
print('depth0_map.shape', depth0_map.shape)
# Get camera pose
kp_all_l = [l.split("\n")[0] for l in open(pose_fn).readlines()][4:]
header0, kp0_l = [],[]
for i, l in enumerate(kp_all_l):
if l.split(" ")[-1] == ('%s.png'%img0_root_fn):
header0 = l
kp0_l = kp_all_l[i+1]
kp0_l = kp0_l.split(" ")
kp0_l = np.reshape(np.array(kp0_l), [-1,3]).astype(float)
q0 = np.array(header0.split(" ")[1:5]).astype(np.float)
t0 = np.array(header0.split(" ")[5:8]).astype(np.float)
R0 = bench_tools.qvec2rotmat(q0)
img0 = cv2.cvtColor(img0, cv2.COLOR_RGB2GRAY)
img0_bw = img0.copy()
img0 = img0[None,...,None].astype(np.float32) / 255.0 # normalize 0-1
assert img0.ndim == 4 # [1,H,W,1]
# Dump keypoint locations and their features
outs = sess.run( {'kpts': ops['kpts'],'feats': ops['feats']},
feed_dict= {photo_ph: img0})
pts0 = outs['kpts'].T
des0 = outs['feats']
kp0 = []
for pt in pts0.T:
kp = cv2.KeyPoint(x=pt[0],y=pt[1], _size=2,
_angle=0, _response=0, _octave=0, _class_id=0)
kp0.append(kp)
kp_on_img0 = np.tile(np.expand_dims(img0_bw,2), (1,1,3))
for i,kp in enumerate(kp0):
pt = (int(round(kp.pt[0])), int(round(kp.pt[1])))
cv2.circle(kp_on_img0, pt, 1, (0, 255, 0), -1, lineType=16)
for img_id in range(1,img_num):
# get img
print('** %04d ** '%img_id)
f.write('** %04d **\n'%img_id)
img1_root_fn = '%04d'%img_id
img1_fn = os.path.join(scene_dir, '%s.png'%img1_root_fn)
img1 = cv2.imread(img1_fn)
img1 = cv2.resize(img1, new_size, interpolation=cv2.INTER_AREA)
# get depth
depth1_fn = '%s/%s.png.photometric.bin'%(scene_dir, img1_root_fn)
depth1_map = bench_tools.read_array(depth1_fn)
depth1_map = cv2.resize(depth1_map, old_size, interpolation=cv2.INTER_CUBIC)
#print('depth1_map.shape', depth1_map.shape)
#raw_input('wait')
# get camera pose
header1, kp1_l = [],[] # camera pose, pixels list
for i, l in enumerate(kp_all_l):
if l.split(" ")[-1] == ('%s.png'%img1_root_fn):
header1 = l
kp1_l = kp_all_l[i+1]
kp1_l = kp1_l.split(" ")
kp1_l = np.reshape(np.array(kp1_l), [-1,3]).astype(float)
q1 = np.array(header1.split(" ")[1:5]).astype(np.float)
t1 = np.array(header1.split(" ")[5:8]).astype(np.float)
R1 = bench_tools.qvec2rotmat(q1)
img1 = cv2.cvtColor(img1, cv2.COLOR_RGB2GRAY)
img1_bw = img1.copy()
img1 = img1[None,...,None].astype(np.float32) / 255.0 # normalize 0-1
assert img1.ndim == 4 # [1,H,W,1]
# Dump keypoint locations and their features
outs = sess.run( {'kpts': ops['kpts'],'feats': ops['feats']},
feed_dict= {photo_ph: img1})
pts1 = outs['kpts'].T
des1 = outs['feats']
kp1 = []
for pt in pts1.T:
kp = cv2.KeyPoint(x=pt[0],y=pt[1], _size=2,
_angle=0, _response=0, _octave=0, _class_id=0)
kp1.append(kp)
kp_on_img1 = np.tile(np.expand_dims(img1_bw,2), (1,1,3))
for i,kp in enumerate(kp1):
pt = (int(round(kp.pt[0])), int(round(kp.pt[1])))
cv2.circle(kp_on_img1, pt, 1, (0, 255, 0), -1, lineType=16)
# scale kp to original size (i.e. the big img) for projection
# f**k floats in python
# let's go
h_pts0 = np.vstack([[sx2big*kp.pt[0], sy2big*kp.pt[1], 1] for kp in kp0])
h_pts1 = np.vstack([[sx2big*kp.pt[0], sy2big*kp.pt[1], 1] for kp in kp1])
rep, N1, N2, M = bench_tools.rep(old_size, h_pts0, h_pts1, K, depth0_map,
depth1_map, R0, t0, R1, t1, cst.THRESH_OVERLAP*sx2big)
print('rep: %.3f - N1: %d - N2: %d - M: %d'%(rep,N1,N2,M))
f.write('rep:%.3f - N1:%d - N2:%d - M:%d\n' %(rep,N1,N2,M))
ms, N1, N2, M_len, M_d_len, inter = bench_tools.ms(old_size, des0,
des1, h_pts0, h_pts1, K, depth0_map, depth1_map, R0, t0, R1,
t1, cst.THRESH_OVERLAP*sx2big, cst.THRESH_DESC, norm='L2')
print('ms:%.3f - N1:%d - N2:%d - M:%d - M_d:%d - inter:%d'
%(ms,N1,N2,M_len, M_d_len, inter))
f.write('ms:%.3f - N1:%d - N2:%d - M:%d - M_d:%d - inter:%d\n'
%(ms, N1, N2, M_len, M_d_len, inter))
if cst.DEBUG:
# match sift
good = []
matches = matcher.knnMatch(des0, des1,k=2)
for i,(m,n) in enumerate(matches):
if m.distance < 0.8*n.distance:
good.append(m)
match_des_img = cv2.drawMatches(img0_bw, kp0, img1_bw, kp1,
good, None, flags=2)
cv2.imshow('match_des', match_des_img)
cv2.imshow('kp_on_img0', np.hstack((kp_on_img0,
kp_on_img1)))
cv2.waitKey(0)
f.close()
print('Done.')
import cv2 as cv
import numpy as np
IMAGE_PATH1 = './images/a.jpg'
IMAGE_PATH2 = './images/c.jpg'
img1 = cv.imread(IMAGE_PATH1)
img2 = cv.imread(IMAGE_PATH2)
key1 = cv.KeyPoint(0, 0, 1)
key2 = cv.KeyPoint(0, 0, 1)
img_matches = np.empty(
(max(img1.shape[0], img2.shape[0]), img1.shape[1] + img2.shape[1], 3),
dtype=np.uint8)
print(img2[199, 132])
print(img1.shape)
print(img2.shape)
print(img_matches.shape)
img_matches[:img1.shape[0], :img1.shape[1]] = img1
img_matches[:img2.shape[0], img1.shape[1]:img1.shape[1] + img2.shape[1]] = img2
cv.circle(img_matches, (1, 100), 10, (0, 0, 255))
cv.circle(img_matches, (img1.shape[0] + 10, 10), 10, (0, 0, 255))
cv.line(img_matches, (10, 10), (img1.shape[0] + 10, 10), (0, 0, 255))
cv.imshow("test", img_matches)
cv.waitKey()
mx,my,mz = [],[],[]
frames = []
idxs = []
while(cap.isOpened()):
#if fc == 0:
ret, frame = cap.read()
#frames.append(frame)
frame0 = frame
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
gray = np.float32(gray)
orb = cv2.ORB_create()
pts = cv2.goodFeaturesToTrack(gray, 3000, 0.01, 7)
kps = [cv2.KeyPoint(x=pt[0][0], y=pt[0][1], _size=20) for pt in pts]
kps, des = orb.compute(frame, kps)
#for pt in pts:
# x,y = pt.ravel()
# cv2.circle(frame,(x,y), 3, (0,255,0), -1)
#if fc == 1:
ret, frame = cap.read()
bat = []
idx1,idx2 = [],[]
frame1 = frame
bf = cv2.BFMatcher(cv2.NORM_HAMMING)
if prev is not None:
matches = bf.knnMatch(des, prev['des'], k =2)
for m,n in matches:
#! /usr/bin/env python
from optparse import OptionParser
import json
from pprint import pprint
import cv2
import re
import numpy as np
import pickle
import random
import csv
import os
class SaveClass:
def __init__(self, votes, keypoints, descriptors, bodypart, hessianThreshold, nOctaves, nOctaveLayers):
self.votes = votes
self.keypoints = keypoints
self.descriptors = descriptors
self.bodypart = bodypart
self.hessianThreshold = hessianThreshold
self.nOctaves = nOctaves
self.nOctaveLayers = nOctaveLayers
if __name__ == '__main__':
parser = OptionParser()
# Read the options
parser.add_option("", "--train-annotation", dest="train_annotation_file", default="", help="frame level training annotation JSON file")
parser.add_option("", "--train-annotation-list", dest="train_annotation_list", default="",help="list of frame level training annotation JSON files")
parser.add_option("", "--project-path", dest="project_dir", default="", help="path containing data directory")
parser.add_option("", "--mh-neighborhood", dest="mh_neighborhood", type="int", default=10,help="distance from mouth hook for a keyppoint to be considered relevant for training")
parser.add_option("", "--positive-training-datafile", dest="train_data_pos", help="File to save the information about the positive training data")
parser.add_option("", "--negative-training-datafile", dest="train_data_neg", help="File to save the information about the negative training data")
parser.add_option("", "--display", dest="display_level", default=0, type="int",help="display intermediate and final results visually, level 5 for all, level 1 for final, level 0 for none")
count = 0
for kp in keypoints:
# print kp.pt, kp.size
print('kp:{}, size:{}'.format(kp.pt, kp.size))
x = int(kp.pt[0])
y = int(kp.pt[1])
size = int(kp.size) >> 1
# extract angle in keypoint sub window
phase_kp = phase[x - size:x + size, y - size:y + size]
# extract gradient magnitude in keypoint subwindow
magnitude_kp = magnitude[x - size:x + size, y - size:y + size]
# create histogram of angle in subwindow BUT only where magnitude of gradients is non zero! Why? Find an
# answer to that question use np.histogram
non_zero_angle = phase_kp[magnitude_kp != 0]
(hist, bins) = np.histogram(non_zero_angle,
bins=[0, 1, 2, 3, 4, 5, 6, 7, 8])
plot_histogram(hist, bins)
descr[count] = hist
return descr
keypoints = [cv2.KeyPoint(15, 15, 11)]
# test for all test images
test = cv2.imread('./images/hog_test/circle.jpg')
descriptor = compute_simple_hog(test, keypoints)
M, mask = cv.findHomography(corners1, corners2, cv.RANSAC, 5.0)
# matchesMask = mask.ravel().tolist()
# Find the Fundamental Matrix
F, mask = cv.findFundamentalMat(corners1, corners2, cv.FM_RANSAC)
# TODO: This has so much noise it returns a rank 3. Refine F.
matchesMask = mask.ravel().tolist()
# Convert corners to keypoints
x1 = corners1[:, :, 0]
y1 = corners1[:, :, 1]
x2 = corners2[:, :, 0]
y2 = corners2[:, :, 1]
kp1 = [cv.KeyPoint(x1[idx], y1[idx], 1) for idx in range(len(x1))]
kp2 = [cv.KeyPoint(x2[idx], y2[idx], 1) for idx in range(len(x2))]
# Visualize corners in image
corners1 = np.int0(corners1)
corners2 = np.int0(corners2)
for i in corners1:
x, y = i.ravel()
cv.circle(img1, (x, y), 3, 255, -1)
for j in corners2:
u, v = j.ravel()
cv.circle(img2, (u, v), 3, 255, -1)
fig1 = plt.figure(figsize=(18, 16), dpi=80, facecolor='w', edgecolor='k')
def to_cv2_kp(self, kp):
# kp is like [batch_idx, y, x, channel]
return cv2.KeyPoint(int(kp[2]/self.sw), int(kp[1]/self.sh), 0)
def to_cv2_kp(kp):
# kp is like [batch_idx, y, x, channel]
return cv2.KeyPoint(kp[2], kp[1], 0)
keypoints_1 = keypoints_1[idx]
keypoints_2 = sift.detect_image(img2) # detect features
idx = np.where(np.sum(keypoints_2.desc, 1) !=
0) # remove zero descriptors from keypoints
keypoints_2 = keypoints_2[idx]
# feature matching
s_time = time.time()
matches = sift.match_images(keypoints_1,
keypoints_2) # matching between 2 keypoints
print("Feature Matching took took {} sec".format(time.time() - s_time))
# transform into cv2 keypoints
cv_kp1 = [
cv2.KeyPoint(x=keypoints_1.x[i], y=keypoints_1.y[i], _size=20)
for i in range(len(keypoints_1.x))
]
cv_kp2 = [
cv2.KeyPoint(x=keypoints_2.x[i], y=keypoints_2.y[i], _size=20)
for i in range(len(keypoints_2.x))
]
# draw the N first matches
N = 50
draw_matches(img1, cv_kp1, img2, cv_kp2, matches[:N])
# draw the features on img 1
# img3 = np.array([])
# img3 = cv2.drawKeypoints(img2, cv_kp2, img3, color=(0, 0, 255))
# plt.imshow(img3), plt.show()
'''
PROCESS FRAME
'''
global vslog
vslog = logging.getLogger('VSLAM')
logging.basicConfig(level=logging.INFO, format='%(levelname)8s %(message)s')
mpl_logger = logging.getLogger('matplotlib')
mpl_logger.setLevel(logging.WARNING)
def _pickle_keypoints(point):
return cv2.KeyPoint, (*point.pt, point.size, point.angle, point.response,
point.octave, point.class_id)
copyreg.pickle(cv2.KeyPoint().__class__, _pickle_keypoints)
def writer(imgnames, masknames, config_dict, queue):
'''
This funtion creates Frame objects which read images from the disk, detects features
and feature descriptors using settings in the config file. It then puts the object into
a multi-process queue.
This function is designed to run in a separate heap and thus takes everything it needs
in form of parameteres and doesn't rely on global variables.
'''
#TILE_KP = config_dict['use_tiling_non_max_supression']
USE_MASKS = config_dict['use_masks']
USE_CLAHE = config_dict['use_clahe']
FEATURE_DETECTOR_TYPE = config_dict['feature_detector_type']
FEATURE_DESCRIPTOR_TYPE = config_dict['feature_descriptor_type']
return canvas
#
# Load Images
# im1 = cv2.imread( '../image/church1.jpg', 0)
# im2 = cv2.imread( '../image/church2.jpg', 0)
im1 = cv2.imread( '../image/a.png')
im2 = cv2.imread( '../image/b.png')
# sift = cv2.xfeatures2d.SIFT_create()
sift = cv2.xfeatures2d.DAISY_create()
step_size = 5
kp = [cv2.KeyPoint(x, y, step_size) for y in range(0, im1.shape[0], step_size)
for x in range(0, im1.shape[1], step_size)]
im1_keypts = cv2.drawKeypoints(im1,kp, None)
im2_keypts = cv2.drawKeypoints(im2,kp, None)
cv2.imshow( 'im1_keypts', im1_keypts)
cv2.imshow( 'im2_keypts', im2_keypts)
cv2.waitKey(0)
startTime = time.time()
dense_feat1 = sift.compute(im1, kp)
dense_feat2 = sift.compute(im2, kp)
print 'time taken for DAISY in sec : ', time.time() - startTime
#
def get_scale_hist(train_data_dir, param):
kp_file_suffix = "-kp-minsc-2.0.h5"
# Check if scale histogram file exists
hist_file_name = train_data_dir + 'scales-histogram-minsc-' + str(
param.dataset.fMinKpSize) + '.h5'
print(hist_file_name)
if not os.path.exists(hist_file_name):
# read all positive keypoint scales
list_png_file = []
# for files in os.listdir(train_data_dir):
# if files.endswith(".png"):
# list_png_file = list_png_file + [files]
list_png_file = list(paths.list_images(train_data_dir))
# all_scales = np.array(dtype=float)
all_scales = []
for png_file in list_png_file:
# print(png_file)
# kp_file_name = train_data_dir + png_file.replace('.png', '_P.mat')
kp_file_name = png_file.replace('.png', '_P.mat')
kp_file_name = png_file.replace('.jpg', '_P.mat')
# print(kp_file_name)
# if a matlab keypointed file exists:
if os.path.exists(kp_file_name):
print("matlab kp file exists")
cur_pos_kp = scipy.io.loadmat(kp_file_name)['TFeatures']
cur_pos_kp = np.asarray(cur_pos_kp, dtype='float')
all_scales += [cur_pos_kp[5, :].flatten()]
# if not
else:
# print("trying to read ascii kp file...")
# look for a Lowe's ASCII keypoint file
# ascci_kp_file = os.path.splitext(png_file)[0]+".sift_bak"
# print(ascci_kp_file)
# cur_pos_kp = readAsciiSiftFile(ascci_kp_file)[0]
kpfilename = os.path.splitext(png_file)[0] + kp_file_suffix
# print(kpfilename)
with h5py.File(kpfilename, "r") as kpfile:
kps = kpfile['valid_keypoints'][()]
cur_pos_kp = [
cv2.KeyPoint(x[0], x[1], x[2], x[3], int(x[4]),
int(x[5])) for x in kps
]
# print(cur_pos_kp)
# print(type(cur_pos_kp))
# extract a 1D array of kp scale values
all_scales += [x.size for x in cur_pos_kp]
# all_scales += [cur_pos_kp[2, :].flatten()]
# print(len(all_scales))
try:
all_scales = np.concatenate(all_scales)
except ValueError as err:
print(err)
pass
# make histogram
hist, bin_edges = np.histogram(all_scales, bins=100)
hist_c = (bin_edges[1:] + bin_edges[:-1]) * 0.5
# save to h5 file
print("Saving hist file...")
with h5py.File(hist_file_name, 'w') as hist_file:
hist_file['all_scales'] = all_scales
hist_file['histogram_bins'] = hist
hist_file['histogram_centers'] = hist_c
# Load from the histogram file
print("Reading hist file...")
with h5py.File(hist_file_name, 'r') as hist_file:
scale_hist = np.asarray(hist_file['histogram_bins'],
dtype=float).flatten()
# print(scale_hist)
scale_hist /= np.sum(scale_hist)
# print(scale_hist)
scale_hist_c = np.asarray(hist_file['histogram_centers']).flatten()
return scale_hist, scale_hist_c
def harris_detector(input_img, output_img, n_keypoints, block_size, ksize,
scales, threshold):
# Get grayscale of the image
gray = cv.cvtColor(input_img, cv.COLOR_BGR2GRAY)
blured_img = cv.GaussianBlur(gray, ksize=(0, 0), sigmaX=4.5)
# Get gradients over each coordenate
g1 = cv.Sobel(blured_img, -1, 1, 0)
g2 = cv.Sobel(blured_img, -1, 0, 1)
# Compute the gaussian piramid of each gradient and the image.
piramide_g1 = [g1] + piramide_gaussiana(g1, 1, scales)
piramide_g2 = [g2] + piramide_gaussiana(g2, 1, scales)
piramide = [gray] + piramide_gaussiana(gray, 1, scales)
# K is going to store all the keypoints.
k = []
# This is going to be the vector of images to return.
imgs = []
# For each of the scales.
for scale in range(scales):
# Get the gradient directions
dx = piramide_g1[scale]
dy = piramide_g2[scale]
# Get the eigenvalues of the corresponding scale of the piramid.
data = cv.cornerEigenValsAndVecs(piramide[scale], block_size, ksize)
e1 = data[:, :, 0]
e2 = data[:, :, 1]
# Calculate matrix of harmonic means
h_means = harmonic_mean(e1, e2)
h_means[np.isnan(h_means)] = 0
# Calculate local maxima
peaks = peak_local_max(h_means,
min_distance=block_size // 2,
num_peaks=n_keypoints,
threshold_abs=threshold)
print("\tMáximos en la escala ", scale, ": ", len(peaks))
size = (scales - scale + 1) * block_size
# Auxiliar array with the keypoints of the current scale.
k_aux = []
# For each of the local peaks.
for peak in peaks:
# Retrieve local maxima coordenates.
x = peak[0]
y = peak[1]
# Calculate the angle of the gradient in that point.
norm = np.sqrt(dx[x][y] * dx[x][y] + dy[x][y] * dy[x][y])
sin = dy[x][y] / norm if norm > 0 else 0
cos = dx[x][y] / norm if norm > 0 else 0
angle = np.degrees(np.arctan2(sin, cos)) + 180
# Add Keypoints to current vector.
k_aux += [cv.KeyPoint(y, x, _size=size, _angle=angle)]
# Add keypoints to global keypoints vector with the transformed coordenates.
k += [
cv.KeyPoint(y * (2**scale),
x * (2**scale),
_size=size,
_angle=angle)
]
# Copy current scale
scale_img = np.copy(piramide[scale])
# Draw keypoints over it
cv.drawKeypoints(piramide[scale], k_aux, scale_img, 0, 5)
# Add it to the return array
imgs += [scale_img]
# Add all keypoints over the original color image and add it to the return array.
imgs = [cv.drawKeypoints(input_img, k, output_img, flags=5)] + imgs
return imgs, k
def get_ransac_image_byte(img_1, locations_1, descriptors_1, img_2, locations_2, descriptors_2, save_path=None, use_opencv_match_vis=True):
"""
Args:
img_1: image bytes. JPEG, PNG
img_2: image bytes. JPEG, PNG
Return:
ransacn result PNG image as byte
score: number of matching inlier
"""
# Convert image byte to 3 channel numpy array
with Image.open(io.BytesIO(img_1)) as img:
img_1 = load_image_into_numpy_array(img)
with Image.open(io.BytesIO(img_2)) as img:
img_2 = load_image_into_numpy_array(img)
inliers, locations_1_to_use, locations_2_to_use = get_inliers(locations_1, descriptors_1, locations_2, descriptors_2)
# Visualize correspondences, and save to file.
fig, ax = plt.subplots(figsize=IMAGE_SIZE)
inlier_idxs = np.nonzero(inliers)[0]
score = sum(inliers)
if score is None:
score = 0
# # For different size of image, transform img_1 to fit to img_2
# print('img_1 shape', img_1.shape)
# print('img_1 type', type(img_1))
# print('img_2 shape', img_2.shape)
# ratio = float(img_2.shape[1]) / img_1.shape[1]
# print('ratio', ratio)
# resize_img_1 = imresize(img_1, ratio, interp='bilinear', mode=None)
# print('resize_img_1 shape', resize_img_1.shape)
if use_opencv_match_vis:
inlier_matches = []
for idx in inlier_idxs:
inlier_matches.append(cv2.DMatch(idx, idx, 0))
kp1 =[]
for point in locations_1_to_use:
kp = cv2.KeyPoint(point[1], point[0], _size=1)
kp1.append(kp)
kp2 =[]
for point in locations_2_to_use:
kp = cv2.KeyPoint(point[1], point[0], _size=1)
kp2.append(kp)
ransac_img = cv2.drawMatches(img_1, kp1, img_2, kp2, inlier_matches, None, flags=0)
ransac_img = cv2.cvtColor(ransac_img, cv2.COLOR_BGR2RGB)
image_byte = cv2.imencode('.png', ransac_img)[1].tostring()
else:
plot_matches(
ax,
img_1,
img_2,
locations_1_to_use,
locations_2_to_use,
np.column_stack((inlier_idxs, inlier_idxs)),
matches_color='b')
ax.axis('off')
extent = ax.get_window_extent().transformed(fig.dpi_scale_trans.inverted())
buf = io.BytesIO()
fig.savefig(buf, bbox_inches=extent, format='png')
plt.close('all') # close resources.
image_byte = buf.getvalue()
return image_byte, score
def auto_correspondences(im1, im2, face1, face2):
#Good Features to Track detection + SIFT descriptor
im1Gray = cv.cvtColor(im1, cv.COLOR_BGR2GRAY)
im2Gray = cv.cvtColor(im2, cv.COLOR_BGR2GRAY)
# im1Hue = cv.cvtColor(im1, cv.COLOR_BGR2HSV)
# im2Hue = cv.cvtColor(im2, cv.COLOR_BGR2HSV)
# im1Gray = im1Hue[..., 0]
# im2Gray = im2Hue[..., 0]
im1Gray = cv.GaussianBlur(im1Gray, (5, 5), 0.1)
im2Gray = cv.GaussianBlur(im2Gray, (5, 5), 0.1)
x1, y1, w1, h1 = face1[0, 1], face1[0, 1], face1[0, 2], face1[0, 3]
x2, y2, w2, h2 = face2[0, 0], face2[0, 1], face2[0, 2], face2[0, 3]
corners1 = cv.goodFeaturesToTrack(im1Gray[y1:y1 + h1, x1:x1 + w1],
maxCorners=200,
qualityLevel=0.05,
minDistance=9)
# corners1 = np.int0(corners1)
kcorners1 = []
corners2 = cv.goodFeaturesToTrack(im2Gray[y2:y2 + h2, x2:x2 + w2],
maxCorners=200,
qualityLevel=0.05,
minDistance=9)
# corners2 = np.int0(corners2)
kcorners2 = []
for i in corners1:
x, y = i.ravel() + face1[0, :2]
kcorners1.append(cv.KeyPoint(x, y, 32))
for i in corners2:
x, y = i.ravel() + face2[0, :2]
# pt = cv.Point2f(x, y)
kcorners2.append(cv.KeyPoint(x, y, 32))
for i in corners1:
x, y = (i.ravel() + face1[0, :2]).astype(int)
# cv.circle(im1,(x,y),3,255,-1)
# cv.imshow('kp1', im1)
for i in corners2:
x, y = (i.ravel() + face2[0, :2]).astype(int)
# cv.circle(im2,(x,y),3,255,-1)
# cv.imshow('kp2', im2)
sift = cv.xfeatures2d.SIFT_create(nfeatures=4)
# kcorners1, des1 = sift.detectAndCompute(im1Gray, None)
# kcorners2, des2 = sift.detectAndCompute(im2Gray, None)
kcorners1, des1 = sift.compute(im1Gray, kcorners1)
kcorners2, des2 = sift.compute(im2Gray, kcorners2)
# plt.imshow(im3,),plt.show()
# im1c=cv.drawKeypoints(im1,kcorners1,outImage = None, flags=cv.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS)
# im2c=cv.drawKeypoints(im2,kcorners2,outImage = None, flags=cv.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS)
# cv.imshow('kp1', im1c)
# cv.imshow('kp2', im2c)
bf = cv.BFMatcher()
matchesF = bf.knnMatch(des1, des2, k=1)
# good = []
# for m,n in matches:
# if m.distance < 0.99*n.distance:
# good.append([m])
#cv2.drawMatchesKnn expects list of lists as matches.
# im3 = np.zeros_like(im1)
# im3 = cv.drawMatchesKnn(im1,kcorners1,im2,kcorners2,matchesF,flags=2, outImg = None)
# cv.imshow('match', im3)
# print(len(matchesF))
m_forward = np.zeros((len(matchesF), 2))
for i in range(len(matchesF)):
m_forward[i, 0] = matchesF[i][0].queryIdx
m_forward[i, 1] = matchesF[i][0].trainIdx
# print kcorners2[np.int32(m_forward[i, 1])].pt
matchesB = bf.knnMatch(des2, des1, k=1)
# good = []
# for m,n in matches:
# if m.distance < 0.99*n.distance:
# good.append([m])
#cv2.drawMatchesKnn expects list of lists as matches.
# im3 = np.zeros_like(im1)
# im4 = cv.drawMatchesKnn(im2,kcorners2,im1,kcorners1,matchesB,flags=2, outImg = None)
# cv.imshow('match2', im4)
# print(len(matchesB))
m_backward = np.zeros((len(matchesB), 2))
for i in range(len(matchesB)):
m_backward[i, 0] = matchesB[i][0].trainIdx
# print kcorners2[np.int32(m_backward[i, 0])].pt
m_backward[i, 1] = matchesB[i][0].queryIdx
return kcorners1, kcorners2, m_forward, m_backward, matchesF, matchesB
def detectAndCompute(self, img, mask=None):
assert mask is None, 'mask support not implemented'
kps, desc, score = self.extractor.extract(img)
kps = [cv2.KeyPoint(x, y, s, 0, r) for (x, y, s), r in zip(kps, score)]
return kps, desc
out_fast = cv2.VideoWriter("./fast.avi", fourcc, 24.0, (3840,2160))
out_orb = cv2.VideoWriter("./orb.avi", fourcc, 24.0, (3840,2160))
while(True):
ret, frame = video_capture.read()
if(frame is None): break #check for empty frames (en of video)
frame_gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
##Harris
mask_harris = cv2.cornerHarris(np.float32(frame_gray), blockSize=2, ksize=3, k=0.04) #2, 3, 0.04 // 2, 5, 0.07
#mask_harris = cv2.dilate(mask_harris, None)
cutout = np.sort(mask_harris.flatten())[-500] #sort from smaller to higher, then take index for cutout
corners = np.where(mask_harris > cutout)
corners = zip(corners[0], corners[1])
kp = list()
for i in corners: kp.append(cv2.KeyPoint(i[1], i[0], 20))
frame_harris = cv2.drawKeypoints(frame_gray, kp, None, [0, 0, 255],
flags=cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS)
##Shi-Tomasi
#maxCorners: Maximum number of corners to return
#qualityLevel: Parameter characterizing the minimal accepted quality of image corners.
#minDistance: Minimum possible Euclidean distance between the returned corners.
#blockSize: Size of an average block for computing a derivative covariation matrix over each pixel neighborhood.
corners = cv2.goodFeaturesToTrack(frame_gray, maxCorners=500, qualityLevel=0.01, minDistance=10, blockSize=2)
corners = np.int0(corners)
kp = list()
for i in corners: kp.append(cv2.KeyPoint(i.ravel()[0], i.ravel()[1], 20))
frame_shitomasi = cv2.drawKeypoints(frame_gray, kp, None, [0, 0, 255],
blobframe = cv2.drawKeypoints(gray3, keypoints, np.array([]), (0, 0, 255),
cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS)
cv2.imshow("preview", gray2)
cv2.imshow("preview2", frame2)
if blobframe is not None:
cv2.imshow("previewblob", blobframe)
if record and face is not None and center is not None and blobframe is not None:
(x, y, w, h) = face
(px, py) = center
if len(keypoints) > 4:
keypoints = keypoints[0:4]
else:
keypoints = keypoints + (
4 - len(keypoints)) * [cv2.KeyPoint(2, -1, -1)]
print(','.join(
[str(x), str(y), str(px), str(py)] +
[str(x.pt[0])
for x in keypoints] + [str(x.pt[1]) for x in keypoints]))
nf = nf + 1
if time() - ptime > 5:
# print(str(nf/(time()-ptime)))
ptime = time()
nf = 0
rval, frame = vc.read()
rval2, frame2 = vc2.read()
key = cv2.waitKey(20)
if key != -1:
# print(key)
def unpickle_cv2(arr):
index = []
for point in arr:
temp = cv2.KeyPoint(x=point[0][0],y=point[0][1],_size=point[1], _angle=point[2], _response=point[3], _octave=point[4], _class_id=point[5])
index.append(temp)
return np.array(index)
def to_cv2_kp(self, kp):
return cv2.KeyPoint(kp[1], kp[0], kp[2], kp[3] / np.pi * 180)
def set_blur_range(img,
instance_scores,
keypoint_scores,
keypoint_coords,
min_pose_score=0.5,
min_part_score=0.5):
out_img = img
adjacent_keypoints = []
cv_keypoints = []
people_count = []
list_start_blur_y = []
list_end_blur_y = []
list_start_blur_x = []
list_end_blur_x = []
for ii, score in enumerate(instance_scores):
if score < min_pose_score:
continue
new_keypoints = posenet.utils.get_adjacent_keypoints(
keypoint_scores[ii, :], keypoint_coords[ii, :, :],
min_part_score)
adjacent_keypoints.extend(new_keypoints)
people_count.append(ii)
kc2 = []
#print(ii, "번째 객체")
for ks, kc in zip(keypoint_scores[ii, :5],
keypoint_coords[ii, :5, :]):
if ks < min_part_score:
continue
cv_keypoints.append(cv2.KeyPoint(kc[1], kc[0], 10. * ks))
if kc[1] < 0:
kc2.append([kc[0], 1])
elif kc[1] > width:
kc2.append([kc[0], width - 3])
elif kc[0] < 0:
kc2.append([1, kc[1]])
elif kc[0] > height:
kc2.append([height - 3, kc[1]])
elif kc[0] < 0 and kc[1] < 0:
kc2.append([1, 1])
elif kc[0] > height and kc[1] > width:
kc2.append([height - 3, width - 3])
else:
kc2.append(kc)
if len(kc2) >= 3:
# 코
nose_y = int(kc2[0][0]) # y축
nose_x = int(kc2[0][1]) # x축
# 오른쪽눈
eye_right_y = int(kc2[1][0])
eye_right_x = int(kc2[1][1])
# 왼쪽눈
eye_left_y = int(kc2[2][0])
eye_left_x = int(kc2[2][1])
#print("> 1.Nose(y:",nose_y,", x:",nose_x,")","2.EyeRight(y:",eye_right_y,", x:",eye_right_x,")",
# "3.EyeLeft(y:",eye_left_y,", x:",eye_left_x,")")
# y값 최솟값, 최댓값, 중앙값 구하기
value_y = [nose_y, eye_right_y, eye_left_y]
min_value_y = min(value_y)
max_value_y = max(value_y)
mean_value_y = int(sum(value_y) / len(value_y))
# x값 최솟값, 최댓값, 중앙값 구하기
value_x = [nose_x, eye_right_x, eye_left_x]
min_value_x = min(value_x)
max_value_x = max(value_x)
mean_value_x = int(sum(value_x) / len(value_x))
# 블러 크기를 키우기 위한 변수 만들기
value_y_gap = abs(max_value_y - min_value_y)
value_x_gap = abs(max_value_x - min_value_x)
# 블러 씌울 부분 변수 선언
start_blur_y = abs(mean_value_y - int(value_y_gap * 1.7))
end_blur_y = mean_value_y + int(value_y_gap * 2)
start_blur_x = abs(min_value_x - value_y_gap)
end_blur_x = max_value_x + value_y_gap
if start_blur_x == end_blur_x:
end_blur_x += 3
if start_blur_y == end_blur_y:
end_blur_y += 3
# 근만씨 동영상 오류 해결중
if abs(eye_right_x - eye_left_x) > 20:
start_blur_y = abs(min_value_y - value_y_gap)
end_blur_y = nose_y + abs(eye_right_x -
eye_left_x) + value_y_gap
start_blur_x = abs(min_value_x - value_x_gap)
end_blur_x = max_value_x + value_x_gap
#print("[",start_blur_y,":",end_blur_y,", ",start_blur_x,":",end_blur_x,"]")
elif len(kc2) >= 2:
# 코
nose_y = int(kc2[0][0]) # y축
nose_x = int(kc2[0][1]) # x축
# 오른쪽눈
eye_right_y = int(kc2[1][0])
eye_right_x = int(kc2[1][1])
#print("> 1.Nose(y:",nose_y,", x:",nose_x,")","2.EyeRight(y:",eye_right_y,", x:",eye_right_x,")")
# y값 최솟값, 최댓값, 중앙값 구하기
value_y = [nose_y, eye_right_y]
min_value_y = min(value_y)
max_value_y = max(value_y)
mean_value_y = int(sum(value_y) / len(value_y))
# x값 최솟값, 최댓값, 중앙값 구하기
value_x = [nose_x, eye_right_x]
min_value_x = min(value_x)
max_value_x = max(value_x)
mean_value_x = int(sum(value_x) / len(value_x))
# 블러 크기를 키우기 위한 변수 만들기
value_y_gap = abs(max_value_y -
int(sum(value_y) / len(value_y)))
value_x_gap = abs(max_value_x -
int(sum(value_x) / len(value_x)))
# 블러 씌울 부분 변수 선언
start_blur_y = abs(mean_value_y - int(value_y_gap * 1.7))
end_blur_y = mean_value_y + int(value_y_gap * 2)
start_blur_x = abs(mean_value_x - value_y_gap)
end_blur_x = mean_value_x + value_y_gap
if start_blur_x == end_blur_x:
end_blur_x += 3
if start_blur_y == end_blur_y:
end_blur_y += 3
#print("[",start_blur_y,":",end_blur_y,", ",start_blur_x,":",end_blur_x,"]")
else:
# 블러 씌울 부분 변수 선언
start_blur_y = 0
end_blur_y = 0
start_blur_x = 0
end_blur_x = 0
list_start_blur_y.append(start_blur_y)
list_end_blur_y.append(end_blur_y)
list_start_blur_x.append(start_blur_x)
list_end_blur_x.append(end_blur_x)
return list_start_blur_y, list_end_blur_y, list_start_blur_x, list_end_blur_x, people_count
import skimage.data as skid
import cv2
import pylab as plt
import scipy.misc
img = scipy.misc.face()
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
plt.figure(figsize=(20, 10))
plt.imshow(img)
plt.show()
sift = cv2.xfeatures2d.SIFT_create()
step_size = 5
kp = [
cv2.KeyPoint(x, y, step_size) for y in range(0, gray.shape[0], step_size)
for x in range(0, gray.shape[1], step_size)
]
img = cv2.drawKeypoints(gray, kp, img)
plt.figure(figsize=(20, 10))
plt.imshow(img)
plt.show()
dense_feat = sift.compute(gray, kp)
do_motion_fade = False
if do_motion_fade:
motion_mask = motion2(new_frame, base)
else:
motion_mask = None
accum = overlay(new_frame, base, motion_mask)
final = accum
if args.draw_keypoints:
new_filtered = []
if affine_new != None:
affine_T = affine_new.T
for kp in filtered:
a = np.array([kp.pt[0], kp.pt[1], 1.0])
pt = np.dot(a, affine_T)
new_kp = cv2.KeyPoint(pt[0], pt[1], kp.size)
new_filtered.append(new_kp)
final = cv2.drawKeypoints(accum,
new_filtered,
color=(0, 255, 0),
flags=0)
cv2.imshow('bgr', res1)
cv2.imshow('smooth', new_frame)
cv2.imshow('final', final)
#output.write(res1)
if args.write_smooth:
output.write(final)
if 0xFF & cv2.waitKey(5) == 27:
break
continue # goto next scene
if args.save2txt:
out_dir = os.path.join(res_dir, scene_name)
if not os.path.exists(out_dir):
os.makedirs(out_dir)
pts_fn = os.path.join(out_dir, '%d_kp.txt' % 1)
np.savetxt(pts_fn, pts0)
# convert to cv2 kp for prototype homogeneity
kp0 = []
for pt in pts0.T:
kp = cv2.KeyPoint(x=pt[0],
y=pt[1],
_size=4,
_angle=0,
_response=0,
_octave=0,
_class_id=0)
kp0.append(kp)
# draw kp on img
img0 = cv2.cvtColor(img0, cv2.COLOR_BGR2GRAY)
kp_on_img0 = np.tile(np.expand_dims(img0, 2), (1, 1, 3))
for i, kp in enumerate(kp0):
pt = (int(round(kp.pt[0])), int(round(kp.pt[1])))
cv2.circle(kp_on_img0, pt, 1, (0, 255, 0), -1, lineType=16)
# description
des_coarse = sess.run(feats_op[args.feat_name],
feed_dict={img_op: patch})[0, :, :, :]
# 匹配对筛选
min_dist = 1000
max_dist = 0
disdif_avg = 0
# 统计平均距离差
for m, n in matches:
disdif_avg += n.distance - m.distance
disdif_avg = disdif_avg / len(matches)
# print('disdif_avg:', disdif_avg)
for m, n in matches:
#自适应阈值
if n.distance > m.distance + 1*disdif_avg:
# if m.distance < 0.9 * n.distance:
goodMatch.append(m)
p2 = cv2.KeyPoint(kps_right[m.trainIdx][0], kps_right[m.trainIdx][1], 1)
p1 = cv2.KeyPoint(kps_left[m.queryIdx][0], kps_left[m.queryIdx][1], 1)
locations_1_to_use.append([p1.pt[0], p1.pt[1]])
locations_2_to_use.append([p2.pt[0], p2.pt[1]])
dis.append([n.distance, m.distance])
#goodMatch = sorted(goodMatch, key=lambda x: x.distance)
# print('match num is %d' % len(goodMatch))
locations_1_to_use = np.array(locations_1_to_use)
locations_2_to_use = np.array(locations_2_to_use)
dis = np.array(dis)
# Perform geometric verification using RANSAC.
_, inliers = measure.ransac((locations_1_to_use, locations_2_to_use),
transform.AffineTransform,
min_samples=3,
residual_threshold=_RESIDUAL_THRESHOLD,