Evaluation with sysu metric Key: for each query identity, its gallery images from the same camera view are discarded. "Following the original setting in ite dataset"
(distmat, q_pids, g_pids, q_camids, g_camids, max_rank = 20)
| 4 | import pdb |
| 5 | |
| 6 | def eval_sysu(distmat, q_pids, g_pids, q_camids, g_camids, max_rank = 20): |
| 7 | """Evaluation with sysu metric |
| 8 | Key: for each query identity, its gallery images from the same camera view are discarded. "Following the original setting in ite dataset" |
| 9 | """ |
| 10 | num_q, num_g = distmat.shape |
| 11 | if num_g < max_rank: |
| 12 | max_rank = num_g |
| 13 | print("Note: number of gallery samples is quite small, got {}".format(num_g)) |
| 14 | indices = np.argsort(distmat, axis=1) |
| 15 | pred_label = g_pids[indices] |
| 16 | matches = (g_pids[indices] == q_pids[:, np.newaxis]).astype(np.int32) |
| 17 | |
| 18 | # compute cmc curve for each query |
| 19 | new_all_cmc = [] |
| 20 | all_cmc = [] |
| 21 | all_AP = [] |
| 22 | all_INP = [] |
| 23 | num_valid_q = 0. # number of valid query |
| 24 | for q_idx in range(num_q): |
| 25 | # get query pid and camid |
| 26 | q_pid = q_pids[q_idx] |
| 27 | q_camid = q_camids[q_idx] |
| 28 | |
| 29 | # remove gallery samples that have the same pid and camid with query |
| 30 | order = indices[q_idx] |
| 31 | remove = (q_camid == 3) & (g_camids[order] == 2) |
| 32 | keep = np.invert(remove) |
| 33 | |
| 34 | # compute cmc curve |
| 35 | # the cmc calculation is different from standard protocol |
| 36 | # we follow the protocol of the author's released code |
| 37 | new_cmc = pred_label[q_idx][keep] |
| 38 | new_index = np.unique(new_cmc, return_index=True)[1] |
| 39 | new_cmc = [new_cmc[index] for index in sorted(new_index)] |
| 40 | |
| 41 | new_match = (new_cmc == q_pid).astype(np.int32) |
| 42 | new_cmc = new_match.cumsum() |
| 43 | new_all_cmc.append(new_cmc[:max_rank]) |
| 44 | |
| 45 | orig_cmc = matches[q_idx][keep] # binary vector, positions with value 1 are correct matches |
| 46 | if not np.any(orig_cmc): |
| 47 | # this condition is true when query identity does not appear in gallery |
| 48 | continue |
| 49 | |
| 50 | cmc = orig_cmc.cumsum() |
| 51 | |
| 52 | # compute mINP |
| 53 | # refernece Deep Learning for Person Re-identification: A Survey and Outlook |
| 54 | pos_idx = np.where(orig_cmc == 1) |
| 55 | pos_max_idx = np.max(pos_idx) |
| 56 | inp = cmc[pos_max_idx]/ (pos_max_idx + 1.0) |
| 57 | all_INP.append(inp) |
| 58 | |
| 59 | cmc[cmc > 1] = 1 |
| 60 | |
| 61 | all_cmc.append(cmc[:max_rank]) |
| 62 | num_valid_q += 1. |
| 63 |