def write_with_confidence(confidence_bound):
cross_score_path = 'data/top10-test/cross_filter_score.log'
cross_pid_path = 'data/top10-test/cross_filter_pid.log'
filter_pid_path = 'data/top10/conf_filter_pid.log'
filter_score_path = 'data/top10/conf_filter_score.log'
safe_remove(filter_pid_path)
safe_remove(filter_score_path)
score_lines = read_lines(cross_score_path)
persons_scores = [[float(score) for score in score_line.split()] for score_line in score_lines]
max_score = max([max(predict_scores) for predict_scores in persons_scores])
pid_lines = read_lines(cross_pid_path)
write_line_cnt = 0
for i in range(len(persons_scores)):
scores = persons_scores[i]
pids = pid_lines[i].split()
has_write = False
for j in range(len(persons_scores[0])):
confidence = max(scores[j]/max_score, 1 - scores[j]/max_score)
if confidence < confidence_bound:
# print(confidence)
if not has_write:
write_line_cnt += 1
has_write = True
write(filter_pid_path, pids[j] + ' ')
write(filter_score_path, '%f ' % scores[j])
write(filter_pid_path, '\n')
write(filter_score_path, '\n')
return write_line_cnt
def test_fusion(fusion_param, ep=0.5, en=0.01):
# copy sort pickle
safe_remove(fusion_param['distribution_pickle_path'])
try:
# 直接使用训练集的时空模型
shutil.copy(fusion_param['src_distribution_pickle_path'],
fusion_param['distribution_pickle_path'])
print 'copy train track distribute pickle done'
except shutil.Error:
print 'pickle ready'
# merge visual probability and track distribution probability
fusion_st_gallery_ranker(fusion_param)
def store_sorted_deltas(fusion_param):
# 时空模型构建核心函数，
# 存储每对摄像头的时间差分布，
# 共6×6个数组，每个数组长度为该对摄像头统计到的时间差数目
# 这个函数运行前会删除distribution_pickle_path： sorted_deltas.pickle，因此也不会有缓存问题
camera_delta_s = viz_data_for_market(fusion_param)
# 对时间差做排序，在预测的时候能快速定位时间差位置，得到时间差的概率
for camera_delta in camera_delta_s:
for delta_s in camera_delta:
delta_s.sort()
# for python
safe_remove(fusion_param['distribution_pickle_path'])
pickle_save(fusion_param['distribution_pickle_path'], camera_delta_s)
def duke_tracks_realloc(tracks, track_name, data_type):
target_list_path = '../data/%s/%s.list' % (track_name, data_type)
target_track_info_dir_path = '/home/cwh/coding/TrackViz/data/%s' % track_name
safe_mkdir(target_track_info_dir_path)
target_track_dir_path = '/home/cwh/coding/%s' % track_name
safe_mkdir(target_track_dir_path)
target_track_type_dir_path = '/home/cwh/coding/%s/%s' % (track_name,
data_type)
safe_mkdir(target_track_type_dir_path)
for track in tracks:
source_img_name = '%04d_c%d_f%07d.jpg' % (int(
track[0]), track[1], track[2])
target_img_name = '%05d%07d%03d.jpg' % (int(
track[0]), track[2], track[1])
shutil.copyfile(
'/home/cwh/coding/DukeMTMC-reID/%s/%s' %
(data_type, source_img_name),
'%s/%s' % (target_track_type_dir_path, target_img_name))
safe_remove(target_list_path)
def write_unequal_rand_st_model(fusion_param):
# fusion_param = get_fusion_param()
rand_answer_path = fusion_param['answer_path'].replace(ctrl_msg['data_folder_path'],
ctrl_msg['data_folder_path'] + '_uerand')
rand_folder_path = folder(rand_answer_path)
safe_mkdir(rand_folder_path)
# although copy all info including pid info, but not use in later training
shutil.copy(fusion_param['answer_path'], rand_answer_path)
rand_path = rand_folder_path + '/renew_pid.log'
safe_remove(rand_path)
origin_tracks = get_tracks(fusion_param)
pid_cnt = len(origin_tracks)
origin_pids = map(lambda x: x + 1, range(pid_cnt))
persons_rand_predict_idx_s = [random.sample(origin_pids, pid_cnt) for _ in range(pid_cnt)]
viz_pid_path = fusion_param['renew_pid_path']
viz_score_path = fusion_param['renew_ac_path']
viz_pids = read_lines(viz_pid_path)
viz_pids = [per_viz_pids.split() for per_viz_pids in viz_pids]
viz_scores = read_lines(viz_score_path)
viz_scores = [per_viz_scores.split() for per_viz_scores in viz_scores]
viz_same_pids = [
[
int(viz_pid) for viz_pid, viz_score in zip(per_viz_pids, per_viz_scores) if float(viz_score) > 0.7
] for per_viz_scores, per_viz_pids in zip(viz_scores, viz_pids)
]
persons_unequal_rand_predict_idx_s = list()
for i in range(pid_cnt):
diff_persons = list(set(persons_rand_predict_idx_s[i]) ^ set(viz_same_pids[i]))
diff_cnt = len(diff_persons)
persons_unequal_rand_predict_idx_s.append(
random.sample(diff_persons, diff_cnt)
)
write_content = ''
for rand_predict_idx_s in persons_unequal_rand_predict_idx_s:
for rand_predict_idx in rand_predict_idx_s:
write_content += str(rand_predict_idx) + ' '
write_content += '\n'
write(rand_path, write_content)
def duke_tracks_realloc(tracks, track_name, data_type):
target_list_path = '../data/%s/%s.list' % (track_name, data_type)
target_track_info_dir_path = '/home/cwh/coding/TrackViz/data/%s' % track_name
safe_mkdir(target_track_info_dir_path)
target_track_dir_path = '/home/cwh/coding/%s' % track_name
safe_mkdir(target_track_dir_path)
target_track_type_dir_path = '/home/cwh/coding/%s/%s' % (track_name,
data_type)
safe_mkdir(target_track_type_dir_path)
names = list()
for track in tracks:
img_name = '%04d_c%d_f%07d.jpg' % (int(track[0]), track[1], track[2])
shutil.copyfile(
'/home/cwh/coding/DukeMTMC-reID/%s/%s' % (data_type, img_name),
'%s/%s' % (target_track_type_dir_path, img_name))
names.append(img_name)
names.append('\n')
list_stmt = ''.join(names)
safe_remove(target_list_path)
write(target_list_path, list_stmt)
def write_rand_pid(fusion_param):
# 对每张左图, 随机生成250个右图的pid，相当于一个随机的renew_pid.log，rand_path也会每次都删除，所以不存在缓存
# todo 不一定需要生成250个右图
# fusion_param = get_fusion_param()
rand_answer_path = fusion_param['answer_path'].replace(ctrl_msg['data_folder_path'], ctrl_msg['data_folder_path'] + '_rand')
rand_folder_path = folder(rand_answer_path)
safe_mkdir(rand_folder_path)
# although copy all info including pid info, but not use in later training
shutil.copy(fusion_param['answer_path'], rand_answer_path)
rand_path = rand_folder_path + '/renew_pid.log'
safe_remove(rand_path)
origin_tracks = get_tracks(fusion_param)
pid_cnt = len(origin_tracks)
origin_pids = map(lambda x: x+1, range(pid_cnt))
persons_rand_predict_idx_s = [random.sample(origin_pids, pid_cnt) for _ in range(pid_cnt)]
write_content = ''
for rand_predict_idx_s in persons_rand_predict_idx_s:
for rand_predict_idx in rand_predict_idx_s:
write_content += str(rand_predict_idx) + ' '
write_content += '\n'
write(rand_path, write_content)
def simple_fusion_st_gallery_ranker(fusion_param):
ep = fusion_param['ep']
en = fusion_param['en']
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path'] # fusion_param = get_fusion_param()
# answer path is probe path
answer_path = fusion_param['answer_path']
answer_lines = read_lines(answer_path)
query_tracks = list()
for answer in answer_lines:
info = answer.split('_')
if 'bmp' in info[2]:
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
query_tracks.append([info[0], int(info[1][0]), int(info[2])])
elif 'f' in info[2]:
query_tracks.append([info[0], int(info[1][1]), int(info[2][1:-5]), 1])
else:
query_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
gallery_path = fusion_param['gallery_path']
gallery_lines = read_lines(gallery_path)
gallery_tracks = list()
for gallery in gallery_lines:
info = gallery.split('_')
if 'bmp' in info[2]:
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
gallery_tracks.append([info[0], int(info[1][0]), int(info[2])])
elif 'f' in info[2]:
gallery_tracks.append([info[0], int(info[1][1]), int(info[2][1:-5]), 1])
else:
gallery_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
print 'probe and gallery tracks ready'
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
print 'read vision scores and pids ready'
if fusion_param['gt_fusion']:
scale = 6.
else:
scale = 3.
# if 'market_market' in log_path:
if True:
# if 'market_market' in log_path:
for i, person_ap_scores in enumerate(persons_ap_scores):
cur_max_vision = max(person_ap_scores)
cur_min_vision = min(person_ap_scores)
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
persons_ap_scores[i] = np.exp(persons_ap_scores[i] * scale)
cur_max_vision = max(persons_ap_scores[i])
cur_min_vision = min(persons_ap_scores[i])
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
# camera_delta_s = pickle_load('true_market_probe.pck')
print 'load track deltas ready'
smooth = '_grid' in log_path
# smooth = True
persons_track_scores = gallery_track_scores(query_tracks, gallery_tracks, camera_delta_s, fusion_param, smooth=smooth)
print 'track scores ready'
persons_cross_scores = list()
safe_remove(map_score_path)
safe_remove(log_path)
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
cur_track_score = persons_track_scores[i][j]
cross_score = cur_track_score * persons_ap_scores[i][j]
cross_scores.append(cross_score)
if max(cross_scores) == 0:
print 'max_cross_scores %d is 0' % i
persons_cross_scores.append(cross_scores)
print 'fusion scores ready'
max_score_s = [max(predict_cross_scores) for predict_cross_scores in persons_cross_scores]
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if persons_cross_scores[i][j] >= 0:
# diff seq not sort, not rank for max, and not normalize
if max_score_s[i] == 0:
# there exist probe track with same seq, diff camera but value > 1000
print 'max_score_s %d is 0' % i
else:
persons_cross_scores[i][j] /= max_score_s[i]
else:
persons_cross_scores[i][j] *= -1.
print 'fusion scores normalized, diff seq use vision score to rank'
person_score_idx_s = list()
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
print 'sorted scores ready'
safe_remove(log_path)
safe_remove(map_score_path)
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
print 'save sorted fusion scores'
return person_score_idx_s
def fusion_st_gallery_ranker(fusion_param):
ep = fusion_param['ep']
en = fusion_param['en']
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path'] # fusion_param = get_fusion_param()
# answer path is probe path
answer_path = fusion_param['answer_path']
answer_lines = read_lines(answer_path)
query_tracks = list()
for answer in answer_lines:
info = answer.split('_')
if 'bmp' in info[2]:
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
query_tracks.append([info[0], int(info[1][0]), int(info[2])])
elif 'f' in info[2]:
query_tracks.append([info[0], int(info[1][1]), int(info[2][1:-5]), 1])
else:
query_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
gallery_path = fusion_param['gallery_path']
gallery_lines = read_lines(gallery_path)
gallery_tracks = list()
for gallery in gallery_lines:
info = gallery.split('_')
if 'bmp' in info[2]:
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
gallery_tracks.append([info[0], int(info[1][0]), int(info[2])])
elif 'f' in info[2]:
gallery_tracks.append([info[0], int(info[1][1]), int(info[2][1:-5]), 1])
else:
gallery_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
print 'probe and gallery tracks ready'
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
print 'read vision scores and pids ready'
if 'market_market' in log_path:
scale = 10
else:
scale = 3 # 1.5 for direct fusion
if True:
# if 'market_market' in log_path:
for i, person_ap_scores in enumerate(persons_ap_scores):
cur_max_vision = max(person_ap_scores)
cur_min_vision = min(person_ap_scores)
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
persons_ap_scores[i] = np.exp(persons_ap_scores[i] * scale)
cur_max_vision = max(persons_ap_scores[i])
cur_min_vision = min(persons_ap_scores[i])
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
# camera_delta_s = pickle_load('true_market_probe.pck')
print 'load track deltas ready'
rand_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'])
print 'load rand deltas ready'
diff_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'].replace('rand', 'diff'))
print 'load diff deltas ready'
# todo tmp diff deltas
# diff_delta_s = rand_delta_s
rand_track_scores = gallery_track_scores(query_tracks, gallery_tracks, rand_delta_s, fusion_param, smooth=False)
print 'rand scores ready'
smooth = '_grid' in log_path
# smooth = True
persons_track_scores = gallery_track_scores(query_tracks, gallery_tracks, camera_delta_s, fusion_param, smooth=smooth)
print 'track scores ready'
# diff_track_scores = gallery_track_scores(query_tracks, gallery_tracks, diff_delta_s, fusion_param, smooth=smooth)
print 'diff track score ready'
# todo tmp diff scores
diff_track_scores = rand_track_scores
persons_cross_scores = list()
safe_remove(map_score_path)
safe_remove(log_path)
# fusion_track_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
# for i, person_ap_pids in enumerate(persons_ap_pids):
# for j, person_ap_pid in enumerate(person_ap_pids):
# cur_track_score = persons_track_scores[i][j]
# rand_track_score = rand_track_scores[i][j]
# if rand_track_score < 0.00002:
# rand_track_score = 0.00002
# fusion_track_scores[i][j] = (cur_track_score * (1 - ep) - en * diff_track_scores[i][j]) / rand_track_score
# for i, person_ap_pids in enumerate(persons_ap_pids):
# cur_max_predict = max(persons_track_scores[i])
# cur_max_rand = max(rand_track_scores[i])
# for j in range(len(fusion_track_scores[i])):
# if fusion_track_scores[i][j] >= 0:
# fusion_track_scores[i][j] /= cur_max_predict/cur_max_rand
# else:
# fusion_track_scores[i][j] = 1.
min_rand = 1e-3 # 0.00002
# min_rand = 2e-5 # 0.00002
if 'market_market' in log_path:
min_rand = 1e-2 # 0.00002
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
cur_track_score = persons_track_scores[i][j]
rand_track_score = rand_track_scores[i][j]
if rand_track_score < 0:
rand_track_score =min_rand
elif rand_track_score < min_rand:
rand_track_score = min_rand
if cur_track_score != 0:
cur_track_score = -1
cross_score = (cur_track_score * (1 - ep) - en * diff_track_scores[i][j]) * (
persons_ap_scores[i][j] + ep / (1 - ep - en)) / rand_track_score
if cur_track_score > 0 and cross_score < 0:
cross_score = 0
cross_scores.append(cross_score)
if max(cross_scores) == 0:
print 'max cross scores %d is 0' % i
persons_cross_scores.append(cross_scores)
print 'fusion scores ready'
# pickle_save(ctrl_msg['data_folder_path']+'viper_r-testpersons_cross_scores.pick', persons_cross_scores)
# pickle_save(ctrl_msg['data_folder_path']+'viper_r-testpersons_ap_pids.pick', persons_ap_pids)
max_score_s = [max(predict_cross_scores) for predict_cross_scores in persons_cross_scores]
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if persons_cross_scores[i][j] >= 0:
# diff seq not sort, not rank for max, and not normalize
if max_score_s[i] == 0:
# there exist probe track with same seq, diff camera but value > 1000
print 'max_score_s %d is 0' % i
else:
persons_cross_scores[i][j] /= max_score_s[i]
# persons_cross_scores[i][j] /= max_score
# if persons_cross_scores[i][j] > 0.5:
# print 'same'
# print persons_cross_scores[i][j]
else:
# so diff seq is negative, normalize by minimum
# persons_cross_scores[i][j] /= min_score_s[i]
# persons_cross_scores[i][j] *= 1.0
persons_cross_scores[i][j] *= -1 * min_rand
if i == 0 and j % 100 == 0:
print 'track: %f vision: %f rand: %f final: %f' % (
persons_track_scores[i][j], persons_ap_scores[i][j], rand_track_scores[i][j], persons_cross_scores[i][j])
print 'fusion scores normalized, diff seq use vision score to rank'
person_score_idx_s = list()
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
print 'sorted scores ready'
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
print 'save sorted fusion scores'
# for i, person_ap_pids in enumerate(persons_ap_pids):
# for j in range(len(person_ap_pids)):
# write(log_path, '%d ' % person_ap_pids[person_score_idx_s[i][j]])
# write(map_score_path, '%.3f ' % persons_cross_scores[i][person_score_idx_s[i][j]])
# write(log_path, '\n')
# write(map_score_path, '\n')
return person_score_idx_s
def simple_fusion_st_img_ranker(fusion_param):
ep = fusion_param['ep']
en = fusion_param['en']
# 从renew_pid和renew_ac获取预测的人物id和图像分数
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
# 从磁盘获取之前建立的时空模型，以及随机时空模型
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
real_tracks = train_tracks(fusion_param)
# 计算时空评分和随机时空评分
persons_track_scores = predict_track_scores(real_tracks, camera_delta_s, fusion_param)
persons_cross_scores = list()
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path']
safe_remove(map_score_path)
safe_remove(log_path)
line_log_cnt = 10
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
cross_score = persons_track_scores[i][j] * persons_ap_scores[i][j]
cross_scores.append(cross_score)
persons_cross_scores.append(cross_scores)
print 'img score ready'
max_score = max([max(predict_cross_scores) for predict_cross_scores in persons_cross_scores])
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if person_cross_score > 0:
# diff seq not sort and not normalize
persons_cross_scores[i][j] /= max_score
else:
persons_cross_scores[i][j] *= -1.
# if real_tracks[i][1] == real_tracks[persons_ap_pids[i][j]][1]:
# # print '%d, %d' % (i, j)
# persons_cross_scores[i][j] = 0
person_score_idx_s = list()
top1_scores = list()
print 'above person score ready'
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
# 统计top1分布，后面计算中位数用
top1_scores.append(person_cross_scores[sort_score_idx_s[0]])
# 降序排，取前60%处的分数
sorted_top1_scores = sorted(top1_scores, reverse=True)
mid_score = sorted_top1_scores[int(len(sorted_top1_scores) * 0.5)]
mid_score_path = fusion_param['mid_score_path']
safe_remove(mid_score_path)
write(mid_score_path, '%f\n' % mid_score)
print(str(mid_score))
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
def fusion_st_img_ranker(fusion_param):
ep = fusion_param['ep']
en = fusion_param['en']
# 从renew_pid和renew_ac获取预测的人物id和图像分数
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
# 从磁盘获取之前建立的时空模型，以及随机时空模型
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
rand_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'])
# diff_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'].replace('rand', 'diff'))
real_tracks = train_tracks(fusion_param)
# 计算时空评分和随机时空评分
persons_track_scores = predict_track_scores(real_tracks, camera_delta_s, fusion_param)
rand_track_scores = predict_track_scores(real_tracks, rand_delta_s, fusion_param)
diff_track_scores = rand_track_scores #predict_track_scores(real_tracks, diff_delta_s, fusion_param)
persons_cross_scores = list()
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path']
safe_remove(map_score_path)
safe_remove(log_path)
line_log_cnt = 10
for i, person_ap_scores in enumerate(persons_ap_scores):
cur_max_vision = max(person_ap_scores)
cur_min_vision = min(person_ap_scores)
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
persons_ap_scores[i] = np.exp(persons_ap_scores[i] * 3)
cur_max_vision = max(persons_ap_scores[i])
cur_min_vision = min(persons_ap_scores[i])
persons_ap_scores[i] = (persons_ap_scores[i] - cur_min_vision) / (cur_max_vision - cur_min_vision)
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
if rand_track_scores[i][j] < 0.00002:
cross_score = (persons_track_scores[i][j]*(1-ep) - en*diff_track_scores[i][j]) * (persons_ap_scores[i][j]+ep/(1-ep-en)) / 0.00002
else:
cross_score = (persons_track_scores[i][j] * (1 - ep) - en * diff_track_scores[i][j]) * (
persons_ap_scores[i][j] + ep / (1 - ep - en)) / rand_track_scores[i][j]
cross_scores.append(cross_score)
persons_cross_scores.append(cross_scores)
print 'img score ready'
max_score = max([max(predict_cross_scores) for predict_cross_scores in persons_cross_scores])
print 'max_cross_score %f' % max_score
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if person_cross_score > 0: # diff camera
# diff seq not sort and not normalize
persons_cross_scores[i][j] /= max_score
else: # same camera and (diff camere && rand score == -1 )
persons_cross_scores[i][j] = persons_ap_scores[i][j]
# if real_tracks[i][1] == real_tracks[persons_ap_pids[i][j]][1]:
# # print '%d, %d' % (i, j)
# persons_cross_scores[i][j] = 0
if i == 0 and j % 100 == 0:
print 'st:%f st_diff%f vis:%f fus:%f' % (
persons_track_scores[i][j], diff_track_scores[i][j], persons_ap_scores[i][j], persons_cross_scores[i][j])
person_score_idx_s = list()
top1_scores = list()
print 'above person score ready'
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
# 统计top1分布，后面计算中位数用
top1_scores.append(person_cross_scores[sort_score_idx_s[0]])
# 降序排，取前60%处的分数
sorted_top1_scores = sorted(top1_scores, reverse=True)
mid_score = sorted_top1_scores[int(len(sorted_top1_scores) * 0.5)]
mid_score_path = fusion_param['mid_score_path']
safe_remove(mid_score_path)
write(mid_score_path, '%f\n' % mid_score)
print(str(mid_score))
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
def get_predict_delta_tracks(fusion_param, useful_predict_limit=10, random=False, diff_person=False, use_real_st=False):
# 获取左图列表
answer_path = fusion_param['answer_path']
answer_lines = read_lines(answer_path)
camera_cnt = 6
real_tracks = list()
for answer in answer_lines:
info = answer.split('_')
if 'bmp' in info[2]:
#
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
# grid
real_tracks.append([info[0], int(info[1][0]), int(info[2]), 1])
elif 'f' in info[2]:
real_tracks.append([info[0], int(info[1][1]), int(info[2][1:-5]), 1])
camera_cnt = 8
else:
# market
real_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
print 'left image ready'
# 获取右图列表
renew_pid_path = fusion_param['renew_pid_path']
predict_lines = read_lines(renew_pid_path)
print 'predict images ready'
# 左图中的人在右图可能出现在6个摄像头中
camera_delta_s = [[list() for j in range(camera_cnt)] for i in range(camera_cnt)]
person_cnt = len(answer_lines)
# market1501数据集有六个序列，只有同一个序列才能计算delta
if random:
useful_predict_limit = max(len(predict_lines)/100, 10)
for i, line in enumerate(predict_lines):
predict_pids = line.split(' ')
useful_cnt = 0
for j, predict_pid in enumerate(predict_pids):
if useful_cnt > useful_predict_limit:
break
if random:
predict_pid = randint(0, person_cnt - 1)
elif diff_person:
predict_pid = randint(10, person_cnt - 1)
else:
# todo transfer: if predict by python, start from 0, needn't minus 1
predict_pid = int(predict_pid)
predict_pid = int(predict_pid)
# same seq
# todo ignore same camera track
if real_tracks[i][3] == real_tracks[predict_pid][3] and real_tracks[i][1] != real_tracks[predict_pid][1]:
# and pid equal: real st
# if use_real_st and random or real_tracks[i][0] == real_tracks[predict_pid][0]:
if True:
useful_cnt += 1
delta = real_tracks[i][2] - real_tracks[predict_pid][2]
if abs(delta) < 1000000:
camera_delta_s[real_tracks[i][1] - 1][real_tracks[predict_pid][1] - 1].append(delta)
print 'deltas collected'
for camera_delta in camera_delta_s:
for delta_s in camera_delta:
delta_s.sort()
print 'deltas sorted'
# for python
safe_remove(fusion_param['distribution_pickle_path'])
pickle_save(fusion_param['distribution_pickle_path'], camera_delta_s)
print 'deltas saved'
return camera_delta_s
def fusion_st_gallery_ranker(fusion_param):
print 'st_filter_market.fusion'
ep = fusion_param['ep']
en = fusion_param['en']
log_path = fusion_param['eval_fusion_path']
map_score_path = fusion_param['fusion_normal_score_path'] # fusion_param = get_fusion_param()
# answer path is probe path
answer_path = fusion_param['answer_path']
answer_lines = read_lines(answer_path)
query_tracks = list()
for answer in answer_lines:
info = answer.split('_')
if 'bmp' in info[2]:
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
query_tracks.append([info[0], int(info[1][0]), int(info[2])])
else:
query_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
gallery_path = fusion_param['gallery_path']
gallery_lines = read_lines(gallery_path)
gallery_tracks = list()
for gallery in gallery_lines:
info = gallery.split('_')
if 'bmp' in info[2]:
info[2] = info[2].split('.')[0]
if len(info) > 4 and 'jpe' in info[6]:
gallery_tracks.append([info[0], int(info[1][0]), int(info[2])])
else:
gallery_tracks.append([info[0], int(info[1][1]), int(info[2]), int(info[1][3])])
print 'probe and gallery tracks ready'
persons_ap_scores = predict_img_scores(fusion_param)
persons_ap_pids = predict_pids(fusion_param)
print 'read vision scores and pids ready'
for i, person_ap_scores in enumerate(persons_ap_scores):
cur_max_vision = 0
for j, person_ap_score in enumerate(person_ap_scores):
if query_tracks[i][1] != gallery_tracks[persons_ap_pids[i][j]][1]:
# diff vision
cur_max_vision = person_ap_score
persons_ap_scores[i] /= cur_max_vision
camera_delta_s = pickle_load(fusion_param['distribution_pickle_path'])
# camera_delta_s = pickle_load('true_market_probe.pck')
print 'load track deltas ready'
rand_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'])
print 'load rand deltas ready'
# diff_delta_s = pickle_load(fusion_param['rand_distribution_pickle_path'].replace('rand', 'diff'))
# todo tmp diff deltas
diff_delta_s = rand_delta_s
rand_track_scores = gallery_track_scores(query_tracks, gallery_tracks, rand_delta_s, fusion_param)
print 'rand scores ready'
persons_track_scores = gallery_track_scores(query_tracks, gallery_tracks, camera_delta_s, fusion_param)
print 'track scores ready'
# diff_track_scores = gallery_track_scores(diff_delta_s, fusion_param)
# todo tmp diff scores
diff_track_scores = rand_track_scores
persons_cross_scores = list()
safe_remove(map_score_path)
safe_remove(log_path)
# fusion_track_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
# for i, person_ap_pids in enumerate(persons_ap_pids):
# for j, person_ap_pid in enumerate(person_ap_pids):
# cur_track_score = persons_track_scores[i][j]
# rand_track_score = rand_track_scores[i][j]
# if rand_track_score < 0.00002:
# rand_track_score = 0.00002
# fusion_track_scores[i][j] = (cur_track_score * (1 - ep) - en * diff_track_scores[i][j]) / rand_track_score
# for i, person_ap_pids in enumerate(persons_ap_pids):
# cur_max_predict = max(persons_track_scores[i])
# cur_max_rand = max(rand_track_scores[i])
# for j in range(len(fusion_track_scores[i])):
# if fusion_track_scores[i][j] >= 0:
# fusion_track_scores[i][j] /= cur_max_predict/cur_max_rand
# else:
# fusion_track_scores[i][j] = 1.
for i, person_ap_pids in enumerate(persons_ap_pids):
cross_scores = list()
for j, person_ap_pid in enumerate(person_ap_pids):
cur_track_score = persons_track_scores[i][j]
rand_track_score = rand_track_scores[i][j]
if rand_track_score < 0:
rand_track_score = 0.00002
elif rand_track_score < 0.00002:
rand_track_score = 0.00002
cur_track_score = -1
cross_score = (cur_track_score * (1 - ep) - en * diff_track_scores[i][j]) * (
persons_ap_scores[i][j] + ep / (1 - ep - en)) / rand_track_score
cross_scores.append(cross_score)
persons_cross_scores.append(cross_scores)
print 'fusion scores ready'
# pickle_save(ctrl_msg['data_folder_path']+'viper_r-testpersons_cross_scores.pick', persons_cross_scores)
# pickle_save(ctrl_msg['data_folder_path']+'viper_r-testpersons_ap_pids.pick', persons_ap_pids)
max_score_s = [max(predict_cross_scores) for predict_cross_scores in persons_cross_scores]
for i, person_cross_scores in enumerate(persons_cross_scores):
for j, person_cross_score in enumerate(person_cross_scores):
if persons_cross_scores[i][j] >= 0:
# diff seq not sort, not rank for max, and not normalize
if max_score_s[i] == 0:
# there exist probe track with same seq, diff camera but value > 1000
print i
else:
persons_cross_scores[i][j] /= max_score_s[i]
# persons_cross_scores[i][j] /= max_score
# if persons_cross_scores[i][j] > 0.5:
# print 'same'
# print persons_cross_scores[i][j]
else:
# so diff seq is negative, normalize by minimum
# persons_cross_scores[i][j] /= min_score_s[i]
# persons_cross_scores[i][j] *= 1.0
persons_cross_scores[i][j] *= -0.00002
# print persons_cross_scores[i][j]
print 'fusion scores normalized, diff seq use vision score to rank'
person_score_idx_s = list()
for i, person_cross_scores in enumerate(persons_cross_scores):
# 单个probe的预测结果中按score排序，得到index，用于对pid进行排序
sort_score_idx_s = sorted(range(len(person_cross_scores)), key=lambda k: -person_cross_scores[k])
person_score_idx_s.append(sort_score_idx_s)
sorted_persons_ap_pids = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
sorted_persons_ap_scores = np.zeros([len(persons_ap_pids), len(persons_ap_pids[0])])
for i, person_ap_pids in enumerate(persons_ap_pids):
for j in range(len(person_ap_pids)):
sorted_persons_ap_pids[i][j] = persons_ap_pids[i][person_score_idx_s[i][j]]
sorted_persons_ap_scores[i][j] = persons_cross_scores[i][person_score_idx_s[i][j]]
print 'sorted scores ready'
np.savetxt(log_path, sorted_persons_ap_pids, fmt='%d')
np.savetxt(map_score_path, sorted_persons_ap_scores, fmt='%f')
print 'save sorted fusion scores'
# for i, person_ap_pids in enumerate(persons_ap_pids):
# for j in range(len(person_ap_pids)):
# write(log_path, '%d ' % person_ap_pids[person_score_idx_s[i][j]])
# write(map_score_path, '%.3f ' % persons_cross_scores[i][person_score_idx_s[i][j]])
# write(log_path, '\n')
# write(map_score_path, '\n')
return person_score_idx_s
def get_predict_tracks(fusion_param, useful_predict_cnt=10):
# 这个函数根据预测的pid，生成更多的图片名，从而构造时空模型，
# 例如，左图1_c1_t2的匹配图片id是2_c2_t3,3_c3_t3,6_c4_t1,
# 则生成1_c2_t3, 1_c3_t3, 1_c4_t1，用于构造时空概率模型
# useful_predcit_cnt为10,实际上会对每张左图的名字产生10个右图的名字，加上原图就是11个新的名字
# 为了加快后续检索速度，将生成的图片名中，摄像头相同的，写到同一个文件里，即predict_camera_path: predict_c%d.txt
# 每次运行这个函数都会删除predict_c%d.txt和predict_tracks.txt，所以不会有缓存旧结果的情况
# todo: 实际上可以在这一步直接生成cameras_deltas,之前是出于重用可视化代码考虑才使用了delta_track.py中的代码
# renew_pid_path: data/top-m2g-std0-train/renew_pid.log'，包含左图预测的图片id, 250*249
renew_pid_path = fusion_param['renew_pid_path']
# predict_track_path：data/top-m2g-std0-train/predict_tracks.txt，存储get predict tracks结果
predict_track_path = fusion_param['predict_track_path']
# 获取左图列表
origin_tracks = get_tracks(fusion_param)
#
safe_remove(predict_track_path)
camera_cnt = 8
global predict_line_idx
predict_line_idx = 0
for i in range(camera_cnt):
safe_remove(fusion_param['predict_camera_path'] + str(i) + '.txt')
def add_predict_track(line):
global predict_line_idx
# print predict_line_idx
if line == '\n':
predict_line_idx += 1
return
if predict_line_idx >= 248:
print(predict_line_idx)
if origin_tracks[predict_line_idx].startswith('-1'):
tail = origin_tracks[predict_line_idx][2:-1]
else:
tail = origin_tracks[predict_line_idx][4:-1]
if 's' in tail:
s_num = int(tail[4])
else:
s_num = 1
if predict_line_idx == 499:
print(predict_line_idx)
if 'jpe' in tail:
camera = tail[1]
else:
camera = tail[2]
track_time = tail.split('_')[2]
mids = line.split()
# 这里写入的是predict_line_idx，而非原来的person id，保证了无监督无标签
write_line(
predict_track_path,
('%04d_c%ds%d_%d_n.jpg' %
(int(predict_line_idx) + 1, int(camera), s_num, int(track_time))))
write_line(
fusion_param['predict_camera_path'] + str(camera) + '.txt',
('%04d_c%ds%d_%d_n.jpg' %
(int(predict_line_idx) + 1, int(camera), s_num, int(track_time))))
for i, mid in enumerate(mids):
if i >= useful_predict_cnt:
break
write_line(predict_track_path,
('%04d_c%ds%d_%d_n.jpg' %
(int(mid), int(camera), s_num, int(track_time))))
write_line(
fusion_param['predict_camera_path'] + str(camera) + '.txt',
('%04d_c%ds%d_%d_n.jpg' %
(int(mid), int(camera), s_num, int(track_time))))
predict_line_idx += 1
# print('done')
read_lines_and(renew_pid_path, add_predict_track)