def sample_cls():
"""sample query images according to cls of test images with occlusion"""
import numpy as np
query_ori_lst = '/home/chencp/data/market-list/query2.lst'
query_lst_new = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/query_cls_test.lst'
query_lst_new_mask = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/query_cls_test-mask.lst'
test_fuse_lst = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/test_fuse_v2.lst'
query_lst = load_lst(query_ori_lst)
test_lst = load_lst(test_fuse_lst)
query_cls = np.array([ins.class_id for ins in query_lst])
test_cls = np.unique([ins.class_id for ins in test_lst])
print('num of cls in test set with occlusion: {}'.format(len(test_cls)))
fout = open(query_lst_new, 'w')
fout_mask = open(query_lst_new_mask, 'w')
cnt = 0
for cls in test_cls:
ind = np.where(query_cls == cls)[0]
# if np.alen(ind) > 2:
# ind = np.random.choice(ind, replace=False, size=2)
cnt += np.alen(ind)
for i in ind:
ins = query_lst[i]
line = '{}\t{}\t{}\n'.format(ins.index, ins.class_id, ins.img_path)
fout.writelines(line)
# mask info
mask_line = line.replace(rgb_root + 'query', mask_root + 'query_seg')
mask_line = mask_line.replace('.jpg', '.png')
fout_mask.writelines(mask_line)
fout.close()
fout_mask.close()
print('sampled {} images, and save in lst file:\n\t{}'.format(cnt, query_lst_new))
def eval_rank_list(rank_list, query_lst_path, gallery_lst_path):
query_lst = load_lst(query_lst_path)
gallery_lst = load_lst(gallery_lst_path)
gallery_cam_lst = np.array([x.cam_id for x in gallery_lst], dtype=np.int32)
gallery_id_lst = np.array([x.class_id for x in gallery_lst],
dtype=np.int32)
num_query = len(query_lst)
num_gallery = len(gallery_lst)
ap = np.zeros((num_query, )) # average precision
cmc = np.zeros((num_query, num_gallery))
for i in tqdm(range(num_query)):
index = np.arange(num_gallery)
good_flag = np.logical_and((gallery_cam_lst != query_lst[i].cam_id),
(gallery_id_lst == query_lst[i].class_id))
junk_flag_1 = (gallery_id_lst == 0)
junk_flag_2 = np.logical_and((gallery_cam_lst == query_lst[i].cam_id),
(gallery_id_lst == query_lst[i].class_id))
good_index = index[good_flag]
junk_index = index[np.logical_or(junk_flag_1, junk_flag_2)]
sort_index = rank_list[i]
ap[i], cmc[i, :] = compute_ap(good_index, junk_index, sort_index)
map = np.mean(ap)
r1 = np.mean(cmc, axis=0)[0]
r5 = np.mean(np.clip(np.sum(cmc[:, :5], axis=1), 0, 1), axis=0)
print('mAP = %f, r1 precision = %f, r5 precision = %f' % (map, r1, r5))
def sample_clean_img():
"""to sample clean img according images with occlusion in query and test list"""
query_lst_ori = '/home/chencp/data/market-list/query2.lst'
test_lst_ori = '/home/chencp/data/market-list/test2.lst'
query_lst_fuse = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/query_fuse_v2.lst'
test_lst_fuse = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/test_fuse_v2.lst'
query_lst_new = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/query_fuse_verify.lst'
test_lst_new = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/test_fuse_verify.lst'
query_lst_mask_new = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/query_fuse_verify-mask.lst'
test_lst_mask_new = '/home/chencp/dataset/binary-annotation-market1501/lst_dir/test_fuse_verify-mask.lst'
query_lst_ori = load_lst(query_lst_ori)
test_lst_ori = load_lst(test_lst_ori)
query_lst_fuse = load_lst(query_lst_fuse)
test_lst_fuse = load_lst(test_lst_fuse)
# sample test set according to query
fout = open(test_lst_new, 'w')
fout_mask = open(test_lst_mask_new, 'w')
cnt = 0
for ins in query_lst_fuse:
ins_sample = [i for i in test_lst_ori if i.class_id == ins.class_id and i.cam_id != ins.cam_id]
cnt += len(ins_sample)
for i_s in ins_sample:
line = '{}\t{}\t{}\n'.format(i_s.index, i_s.class_id, i_s.img_path)
fout.writelines(line)
# mask info
mask_line = line.replace(rgb_root + 'bounding_box_test', mask_root + 'bounding_box_test_seg')
mask_line = mask_line.replace('.jpg', '.png')
fout_mask.writelines(mask_line)
print('sample {} images from test set according to occlusion query set'.format(cnt))
print('\tand save the list file to {}'.format(test_lst_new))
fout.close()
fout_mask.close()
# sample query set according to test
fout = open(query_lst_new, 'w')
fout_mask = open(query_lst_mask_new, 'w')
cnt = 0
for ins in test_lst_fuse:
ins_sample = [i for i in query_lst_ori if i.class_id == ins.class_id and i.cam_id != ins.cam_id]
cnt += len(ins_sample)
for i_s in ins_sample:
line = '{}\t{}\t{}\n'.format(i_s.index, i_s.class_id, i_s.img_path)
fout.writelines(line)
# mask info
mask_line = line.replace(rgb_root + 'query', mask_root + 'query_seg')
mask_line = mask_line.replace('.jpg', '.png')
fout_mask.writelines(mask_line)
print('sample {} images from query set according to occlusion test set'.format(cnt))
print('\tand save the list file to {}'.format(query_lst_new))
fout.close()
fout_mask.close()
"--crop-size 128 --dataset %s" % (
sys.version[0], prefix, gpu, epoch_idx, query_prefix, gallery_prefix, dataset)
cmd = cmd + ' --data_type %s' % data_type
if force_resize:
cmd = cmd + " --force-resize"
if mask_label:
cmd = cmd + ' --mask_label'
if soft_or_hard:
cmd = cmd + ' --soft_or_hard'
if rgbm_iter:
cmd = cmd + ' --rgbm_iter'
if fea_tag:
cmd = cmd + ' --fea_tag %s' % fea_tag
subprocess.check_call(cmd.split(" "))
assert os.path.exists(query_features_path) and os.path.exists(
gallery_features_path)
query_features = sio.loadmat(query_features_path)["feat"]
gallery_features = sio.loadmat(gallery_features_path)["feat"]
query_lst = load_lst(query_prefix + ".lst")
gallery_lst = load_lst(gallery_prefix + ".lst")
eval_feature_verify(query_features,
gallery_features,
query_lst,
gallery_lst,
metric="cosine")
def tsne_viz_query2gallery(prefix, dataset, viz_range):
if not isinstance(viz_range, (list, tuple, int)):
raise ValueError("viz_range must be list or tuple or int")
if isinstance(viz_range, int):
viz_range = list(range(viz_range))
query_features_path = 'features/%s/query-%s.mat' % (dataset, prefix)
gallery_features_path = "features/%s/gallery-%s.mat" % (dataset, prefix)
query_lst_path = "/mnt/truenas/scratch/chuanchen.luo/data/reid/%s-list/query.lst" % dataset
gallery_lst_path = "/mnt/truenas/scratch/chuanchen.luo/data/reid/%s-list/test.lst" % dataset
query_lst = np.array(load_lst(query_lst_path))
gallery_lst = np.array(load_lst(gallery_lst_path))
query_features = sio.loadmat(query_features_path)["feat"]
gallery_features = sio.loadmat(gallery_features_path)["feat"]
query_features = normalize(query_features)
gallery_features = normalize(gallery_features)
for i in viz_range:
q_feat = query_features[i]
dist = -np.dot(gallery_features, q_feat)
rank_list = np.argsort(dist)[:150]
g_feats = gallery_features[rank_list]
q_record = Record(*query_lst[i])
g_records = [Record(*item) for item in gallery_lst[rank_list]]
same_list = [
i for i in range(g_feats.shape[0])
if q_record.class_id == g_records[i].class_id
]
diff_list = [
i for i in range(g_feats.shape[0])
if q_record.class_id != g_records[i].class_id
]
init_embed = TSNE(n_components=2,
init="pca",
perplexity=10,
metric="cosine").fit_transform(g_feats)
W = np.dot(g_feats, g_feats.T)
W = np.exp(W / TEMPERATURE)
W = W / np.sum(W, axis=1, keepdims=True)
g_feats = np.dot(W, g_feats)
res_embed = TSNE(n_components=2,
init="pca",
perplexity=10,
metric="cosine").fit_transform(g_feats)
plt.figure(0)
plt.scatter(init_embed[same_list, 0],
init_embed[same_list, 1],
label="same")
plt.scatter(init_embed[diff_list, 0],
init_embed[diff_list, 1],
label="diff")
plt.savefig("%d_tsne_orignal.png" % i)
plt.close(0)
plt.figure(1)
plt.scatter(res_embed[same_list, 0],
res_embed[same_list, 1],
label="same")
plt.scatter(res_embed[diff_list, 0],
res_embed[diff_list, 1],
label="diff")
plt.savefig("%d_tsne_transformed_%.2f.png" % (i, TEMPERATURE))
plt.close(1)
print(i)
def tsne_viz_iteration(prefix, dataset, query_id):
query_features_path = 'features/%s/query-%s.mat' % (dataset, prefix)
gallery_features_path = "features/%s/gallery-%s.mat" % (dataset, prefix)
query_lst_path = "/mnt/truenas/scratch/chuanchen.luo/data/reid/%s-list/query.lst" % dataset
gallery_lst_path = "/mnt/truenas/scratch/chuanchen.luo/data/reid/%s-list/test.lst" % dataset
query_lst = np.array(load_lst(query_lst_path))
gallery_lst = np.array(load_lst(gallery_lst_path))
query_features = sio.loadmat(query_features_path)["feat"]
gallery_features = sio.loadmat(gallery_features_path)["feat"]
query_features = normalize(query_features)
gallery_features = normalize(gallery_features)
q_feat = query_features[query_id]
dist = -np.dot(gallery_features, q_feat)
rank_list = np.argsort(dist)[:150]
g_feats = gallery_features[rank_list]
q_record = Record(*query_lst[query_id])
g_records = [Record(*item) for item in gallery_lst[rank_list]]
same_list = [
i for i in range(g_feats.shape[0])
if q_record.class_id == g_records[i].class_id
]
diff_list = [
i for i in range(g_feats.shape[0])
if q_record.class_id != g_records[i].class_id
]
init_embed = TSNE(n_components=2,
init="pca",
perplexity=8,
metric="cosine").fit_transform(g_feats)
W = np.dot(g_feats, g_feats.T)
W = np.exp(W / TEMPERATURE)
W = W / np.sum(W, axis=1, keepdims=True)
plt.figure(0)
plt.scatter(init_embed[same_list, 0],
init_embed[same_list, 1],
label="same")
plt.scatter(init_embed[diff_list, 0],
init_embed[diff_list, 1],
label="diff")
plt.savefig("%d_tsne_orignal.png" % query_id)
plt.close(0)
for iteration in range(10):
g_feats = np.dot(W, g_feats)
res_embed = TSNE(n_components=2,
init="pca",
perplexity=8,
metric="cosine").fit_transform(g_feats)
plt.figure(iteration)
plt.scatter(res_embed[same_list, 0],
res_embed[same_list, 1],
label="same")
plt.scatter(res_embed[diff_list, 0],
res_embed[diff_list, 1],
label="diff")
plt.savefig("%d_%02d_tsne_transformed_%.2f.png" %
(query_id, iteration, TEMPERATURE))
plt.close(iteration)
print("%d-%d" % (query_id, iteration))