# experiment 1
# use gardens point datatset to giev the result of MCN's
# unstability when input connections is less
# and the more input connections the more time consumes
# (*) means the params you should specify
# day_right vs. night_right
D1 = np.load('../../desc/netvlad/gp/day_right_desc.npy') #(*)
D2 = np.load('../../desc/netvlad/gp/night_right_desc.npy') #(*)
D1 = D1 / np.linalg.norm(D1, axis=0)
D2 = D2 / np.linalg.norm(D2, axis=0)
err = 3 #(*)
GT = utils.getGroundTruthMatrix(D1.shape[0], err)
nConPerCol = [100, 300, 500, 700, 1000, 1500, 2000] #(*)
# traverse all nConPerCol and compute the AP and time consumption
# and save dict which has key: ap and time
# if the result is saved, you can comment the following code
# params = MCN.MCNParams(probAddCon=0.05,
# nCellPerCol=32,
# nConPerCol=200,
# minColActivity=0.7,
# nColPerPattern=50,
# kActiveCol=100)
#
# MCN_time_cost = []
# MCN_ap = []
#
# 证明环视数据集的优越性，通过对环视描述子压缩，并和原维度单视图描述子比较AP
# 每行打(*)号的需要根据实际情况配置
# 需要omni-SCUT数据集的描述子文件.npy
# ---------------------------- 说明 ----------------------------------
D1 = np.load('./netvlad/day_desc.npy') #(*)
D2 = np.load('./netvlad/dawn_desc.npy') #(*)
D1 = D1 / np.linalg.norm(D1, axis=0)
D2 = D2 / np.linalg.norm(D2, axis=0)
num = D1.shape[0]
omni_D1 = D1.reshape((num // 4, -1))
omni_D2 = D2.reshape((num // 4, -1))
GT = utils.getGroundTruthMatrix(num // 4, 1)
sig_D1 = D1[::4, :]
sig_D2 = D2[::4, :]
del D1, D2
# single view
num, old_dims = sig_D1.shape
new_dims = 8192 # acturally 8192 * 2
s = 0.25
P = np.random.rand(old_dims, new_dims)
P /= np.linalg.norm(P, axis=1, keepdims=True)
sig_D1_slsbh = utils.getLSBH(sig_D1, P, 0.25)
sig_D2_slsbh = utils.getLSBH(sig_D2, P, 0.25)
# if use SCUT datasets, uncomment following code dbFeat = np.load(db_save_path)[::4, :] #(*) qFeat = np.load(q_save_path)[::4, :] # otherwise uncomment following code # dbFeat = np.load(db_save_path) # qFeat = np.load(q_save_path) dbn = dbFeat.shape[0] qn = qFeat.shape[0] # get ground truth #(*) # if not use oxford robotcar, uncomment following code err = arg_dataset['err'] gt = utils.getGroundTruth(dbn, err) gtm = utils.getGroundTruthMatrix(dbn, err) # oxford robotcar # err = arg_dataset['err'] # db_gps = np.load(join(imgdir, 'gps_' + subdir[arg_eval['compare_subdir'][0]] + '.npy')) # q_gps = np.load(join(imgdir, 'gps_' + subdir[arg_eval['compare_subdir'][1]] + '.npy')) # _, gtm = utils.getGpsGT(db_gps, q_gps, err) # following code is optional, if you want to see the recall@N # you can uncomment it # _ = utils.getResult(dbFeat, qFeat, gt, gtm) # experiment 3_3 D1 = dbFeat / np.linalg.norm(dbFeat, axis=0) D2 = qFeat / np.linalg.norm(qFeat, axis=0) del dbFeat, qFeat cannum = int(0.02 * D1.shape[0])
plt.figure()
err = 9 #(*) error tolerance for nordland dataset, 9 is set in paper
S_file = '../cp3_5/vis3/nd_sumwin.npz' #(*) NL similarity matrix
S = np.load(S_file)
S_pw = S['S_pw']
S_pwk = S['S_pwk']
S_mcn = S['S_mcn']
S_seq = loadmat('../cp3_4/seqslam/nd_sumwin.mat')['S'] #(*)
tmp = np.isnan(S_seq)
S_seq[tmp] = np.max(S_seq[~tmp])
S_smcn = S['S_smcn']
S_smcntf = S['S_smcntf']
S_dd = S_nd_dd
del S
GT = utils.getGroundTruthMatrix(S_pw.shape[0], err)
P, R = utils.drawPR(S_pw, GT)
ap = utils.calAvgPred(P, R)
plt.plot(R, P, color=vis.color[0], label='PW: %.4f' % ap, linewidth=2)
P, R = utils.drawPR(S_pwk, GT)
ap = utils.calAvgPred(P, R)
plt.plot(R, P, color=vis.color[1], label='PWk: %.4f' % ap, linewidth=2)
P, R = utils.drawPR(S_mcn, GT)
ap = utils.calAvgPred(P, R)
plt.plot(R, P, color=vis.color[2], label='MCN: %.4f' % ap, linewidth=2)
P, R = utils.drawPR(S_seq, GT, True)
ap = utils.calAvgPred(P, R)
plt.plot(R, P, color=vis.color[3], label='SeqSLAM: %.4f' % ap, linewidth=2)
if root.endswith('scut/'):
qn = qn // 4
picid = sample(range(0, qn), numPicToShow)
numpic = len(picid)
img_format = 'Image%03d.jpg' #(*)
err = 3 #(*)
saveName = './vis2/gp.png' #(*)
if root.endswith('robotcar/'):
db_gps = np.load(root + 'gps_snow.npy')
q_gps = np.load(root + 'gps_night.npy')
_, gt = utils.getGpsGT(db_gps, q_gps, err)
else:
gt = utils.getGroundTruthMatrix(qn, err)
gtl = []
for each in picid:
gtl.append(list(np.where(gt[:, each])[0]))
id_pw = np.argmax(S_pw[:, picid], axis=0)
id_dd = np.argmax(S_dd[:, picid], axis=0)
id_mcn = np.argmax(S_mcn[:, picid], axis=0)
id_smcn = np.argmax(S_smcn[:, picid], axis=0)
with open('./vis3/pathid.pkl', 'rb') as f: #(*)
id_smcntf = pickle.load(f)['gp']
id_smcntf = [id_smcntf[x] for x in picid]
id_seq = np.argmin(S_seq[:, picid], axis=0)
real_id_pw = np.copy(id_pw)
# load features
subdir = arg_dataset['subdir']
db_save_path = join(desc_dir,
subdir[arg_eval['compare_subdir'][0]] + '_desc.npy')
q_save_path = join(desc_dir,
subdir[arg_eval['compare_subdir'][1]] + '_desc.npy')
dbFeat = np.load(db_save_path)
qFeat = np.load(q_save_path)
num, dim = dbFeat.shape
# get ground truth
err = 3
gt = utils.getGroundTruth(num, err)
gtm = utils.getGroundTruthMatrix(num, err)
_ = utils.getResult(dbFeat, qFeat, gt, gtm)
S = utils.MCN_pairwise(dbFeat, qFeat)
P, R = utils.drawPR(S, gtm)
auc = utils.calAvgPred(P, R)
print("auc : %.4f" % auc)
# test dimension reduction effect
# r_dim = 128
# w_dbFeat = utils.pca_whitening(dbFeat, r_dim)
# w_dbFeat = np.ascontiguousarray(w_dbFeat)
# w_qFeat = utils.pca_whitening(qFeat, r_dim)
# w_qFeat = np.ascontiguousarray(w_qFeat)
img_format = '%d.png' #(*)
# name as <dataset name>.png
saveName = './vis3/' + root.split('/')[-2] + '.png' #(*)
ref_saveName = './vis3/ref_' + root.split('/')[-2] + '.png' #(*)
# err tolerance
green_tolerance = 5 #(*)
blue_tolerance = 2 * green_tolerance #(*)
if root.endswith('robotcar/'):
db_gps = np.load(root + 'gps_snow.npy')
q_gps = np.load(root + 'gps_night.npy')
_, gtg = utils.getGpsGT(db_gps, q_gps, green_tolerance)
_, gtb = utils.getGpsGT(db_gps, q_gps, blue_tolerance)
else:
gtg = utils.getGroundTruthMatrix(qn, green_tolerance)
gtb = utils.getGroundTruthMatrix(qn, blue_tolerance)
gtlg = list(np.where(gtg[:, numTopick])[0])
gtlb = list(np.where(gtb[:, numTopick])[0])
# color
blue = np.array([[[255, 0, 0]]])
green = np.array([[[0, 255, 0]]])
red = np.array([[[0, 0, 255]]])
id_pw = np.argsort(-S_pw[:, numTopick])[:topPicToShow]
id_mcn = np.argsort(-S_mcn[:, numTopick])[:topPicToShow]
id_dd = np.argsort(-S_dd[:, numTopick])[:topPicToShow]
id_smcn = np.argsort(-S_smcn[:, numTopick])[:topPicToShow]
id_smcntf = np.argsort(S_smcntf[:, numTopick])[:topPicToShow]
# load id file
with open('./id_dict.pkl', 'rb') as f:
id_dict = pickle.load(f)
sig_f_id = id_dict['sig_f_id']
omni_f_id = id_dict['onmi_f_id']
sig_t_id = id_dict['sig_t_id']
omni_t_id = id_dict['onmi_t_id']
omni_t_mcn_id = id_dict['onmi_t_mcn_id']
omni_t_smcn_id = id_dict['onmi_t_smcn_id']
omni_t_smcntf_id = id_dict['onmi_t_smcntf_id']
# error tolerance
sc_err = 1
# make groudtruth
sc_gtg = utils.getGroundTruthMatrix(sc_num, sc_err)
sc_gtb = utils.getGroundTruthMatrix(sc_num, 2 * sc_err)
# save video
fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
fps = 1.0
video = cv2.VideoWriter('./omni.mp4', fourcc, fps, (900, 600))
for i in trange(show_num):
# first row
st = 4 * i
img = cv2.imread(sc_qfile + sc_img_f % st)
set_img(img, ex_c, ex_r)
set_img(img, ex_c + img_size + ex_r, ex_r)