ReID基础 | 跨模态reid中常用的固定代码

1. 计算rank、mAP和mINP

"""eval_metrics.py"""

from __future__ import print_function, absolute_import
import numpy as np


def eval_sysu(distmat, q_pids, g_pids, q_camids, g_camids, max_rank=20):
    """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传入的是每个特征向量的乘积的负数，gallery（3804*2048）和query转置（2048*301）的矩阵乘积（3804*301）的负数
    # q_pids表示query的标签，g_pids表示gallery的标签
    # q_camids, g_camids分别表示query和gallery样本的相机标签
    # max_rank，统计gallery可能性最大的前多少个

    num_q, num_g = distmat.shape    # num_q表示query的个数，num_g表示gallery的个数

    if num_g < max_rank:
        max_rank = num_g
        print("Note: number of gallery samples is quite small, got {}".format(num_g))

    # 沿着行从小到大排序，返回该数值的原来的索引号（因为传入的是原值的负数，所以矩阵乘积原值中最大的那个数值，序号最小为0）
    # 返回原值中，每行数值从大到小的索引值
    indices = np.argsort(distmat, axis=1)
    # 得到每行（每个query）可能性从大到小的预测标签
    pred_label = g_pids[indices]
    # 将每行的预测标签和真实标签比较
    matches = (g_pids[indices] == q_pids[:, np.newaxis]).astype(np.int32)   # 扩大真实标签的维度，最终将True/False转换数据类型1/0
    # print('matches:', matches, matches.shape)       # matches是3804*301维度，每个元素是1或者0.
    
    # compute cmc curve for each query
    new_all_cmc = []    # 一种新的方法，存储所有query的[000111..]数组
    all_cmc = []        # 原始的方法，存储所有query的[000111...]数组
    all_AP = []         # 存储所有query的AP
    all_INP = []        # 存储所有query的INP
    num_valid_q = 0.    # number of valid query（有效的query个数）

    # 遍历所有的query样本
    for q_idx in range(num_q):
        # 1. get query pid and camid
        q_pid = q_pids[q_idx]
        q_camid = q_camids[q_idx]

        # 2. 要在不同位置的摄像机之间进行匹配
        # 相机2和相机3在相同的位置，所以相机3的probe图像要跳过相机2的gallery图像
        order = indices[q_idx]  # 找到这个query对应的gallery可能性排序索引
        remove = (q_camid == 3) & (g_camids[order] == 2)    # 同时成立则为True，输出301个True或者False
        # 取反，False->True。得到的keep中，True是可以使用的gallery。
        keep = np.invert(remove)

        # 3. compute cmc curve
        # the cmc calculation is different from standard protocol
        # we follow the protocol of the author's released code

        # 去除重复的预测标签
        new_cmc = pred_label[q_idx][keep]       # 取出这个query行，所有True的预测标签
        new_index = np.unique(new_cmc, return_index=True)[1]  # 将重复的预测标签只保留一个，返回重复标签的第一个索引下标
        new_cmc = [new_cmc[index] for index in sorted(new_index)]

        # new_match从找到正确标签开始全是1，之前全是0
        new_match = (new_cmc == q_pid).astype(np.int32)     # 输出1或者0，1表示与query同ID的预测标签
        new_cmc = new_match.cumsum()        # 依次输出前k个元素累加和（k=1，2...） 0 0 0 1 1 1 ...
        new_all_cmc.append(new_cmc[:max_rank])  # 将该样本的序列添加到所有样本的数组中

        # 原始cmc
        orig_cmc = matches[q_idx][keep] # binary vector, positions with value 1 are correct matches
        if not np.any(orig_cmc):
            # this condition is true when query identity does not appear in gallery
            continue
        cmc = orig_cmc.cumsum()         # 0 0 0 0 1 1 1 1 2 2 2 2

        # 4. compute mINP
        # refernece: Deep Learning for Person Re-identification: A Survey and Outlook
        pos_idx = np.where(orig_cmc == 1)               # 找到正确标签对应的索引
        pos_max_idx = np.max(pos_idx)                   # 找到最大的索引
        inp = cmc[pos_max_idx] / (pos_max_idx + 1.0)    # 计算INP
        all_INP.append(inp)

        # 将序列0 0 1 1 2 2  转换为 0 0 1 1 1 1
        cmc[cmc > 1] = 1
        all_cmc.append(cmc[:max_rank])
        num_valid_q += 1.

        # 5. compute average precision(AP)
        # reference: https://en.wikipedia.org/wiki/Evaluation_measures_(information_retrieval)#Average_precision
        num_rel = orig_cmc.sum()        # 把该行中所有的1求和，得到正确样本的个数
        tmp_cmc = orig_cmc.cumsum()     # 累加，得到0 0 1 1 2 2...

        # 正确标签，在正样本的位置 / 在所有样本中的位置
        tmp_cmc = [x / (i+1.) for i, x in enumerate(tmp_cmc)]   # i是索引，x是数值，成对取出
        tmp_cmc = np.asarray(tmp_cmc) * orig_cmc    # 只保留正确标签的计算结果
        AP = tmp_cmc.sum() / num_rel        # 计算AP
        all_AP.append(AP)

    assert num_valid_q > 0,