11def re_ranking(q_g_dist, q_q_dist, g_g_dist, k1: int = 20, k2: int = 6, lambda_value: float = 0.3):
12 original_dist = np.concatenate(
13 [np.concatenate([q_q_dist, q_g_dist], axis=1),
14 np.concatenate([q_g_dist.T, g_g_dist], axis=1)],
15 axis=0)
16 original_dist = np.power(original_dist, 2).astype(np.float32)
17 original_dist = np.transpose(1. * original_dist / np.max(original_dist, axis=0))
18 V = np.zeros_like(original_dist).astype(np.float32)
19 initial_rank = np.argsort(original_dist).astype(np.int32)
20
21 query_num = q_g_dist.shape[0]
22 gallery_num = q_g_dist.shape[0] + q_g_dist.shape[1]
23 all_num = gallery_num
24
25 for i in range(all_num):
26
27 forward_k_neigh_index = initial_rank[i, :k1 + 1]
28 backward_k_neigh_index = initial_rank[forward_k_neigh_index, :k1 + 1]
29 fi = np.where(backward_k_neigh_index == i)[0]
30 k_reciprocal_index = forward_k_neigh_index[fi]
31 k_reciprocal_expansion_index = k_reciprocal_index
32 for j in range(len(k_reciprocal_index)):
33 candidate = k_reciprocal_index[j]
34 candidate_forward_k_neigh_index = initial_rank[candidate,
35 :int(np.around(k1 / 2.)) + 1]
36 candidate_backward_k_neigh_index = initial_rank[candidate_forward_k_neigh_index,
37 :int(np.around(k1 / 2.)) + 1]
38 fi_candidate = np.where(candidate_backward_k_neigh_index == candidate)[0]
39 candidate_k_reciprocal_index = candidate_forward_k_neigh_index[fi_candidate]
40 if len(np.intersect1d(candidate_k_reciprocal_index, k_reciprocal_index)) > 2. / 3 * len(
41 candidate_k_reciprocal_index):
42 k_reciprocal_expansion_index = np.append(k_reciprocal_expansion_index, candidate_k_reciprocal_index)
43
44 k_reciprocal_expansion_index = np.unique(k_reciprocal_expansion_index)
45 weight = np.exp(-original_dist[i, k_reciprocal_expansion_index])
46 V[i, k_reciprocal_expansion_index] = 1. * weight / np.sum(weight)
47 original_dist = original_dist[:query_num, ]
48 if k2 != 1:
49 V_qe = np.zeros_like(V, dtype=np.float32)
50 for i in range(all_num):
51 V_qe[i, :] = np.mean(V[initial_rank[i, :k2], :], axis=0)
52 V = V_qe
53 del V_qe
54 del initial_rank
55 invIndex = []
56 for i in range(gallery_num):
57 invIndex.append(np.where(V[:, i] != 0)[0])
58
59 jaccard_dist = np.zeros_like(original_dist, dtype=np.float32)
60
61 for i in range(query_num):
62 temp_min = np.zeros(shape=[1, gallery_num], dtype=np.float32)
63 indNonZero = np.where(V[i, :] != 0)[0]
64 indImages = [invIndex[ind] for ind in indNonZero]
65 for j in range(len(indNonZero)):
66 temp_min[0, indImages[j]] = temp_min[0, indImages[j]] + np.minimum(V[i, indNonZero[j]],
67 V[indImages[j], indNonZero[j]])
68 jaccard_dist[i] = 1 - temp_min / (2. - temp_min)
69
70 final_dist = jaccard_dist * (1 - lambda_value) + original_dist * lambda_value
71 del original_dist, V, jaccard_dist
72 final_dist = final_dist[:query_num, query_num:]
73 return final_dist
