asdTools-ReID Rank可视化（2023/10/23更新）

Github - 开源代码及Readme
Blog - 工具介绍

摘要：使用markdown实现ReID的Rank可视化的工具介绍。

2023/10/21更新：修复bug
2023/10/23更新：优化单项展示及汇总展示

场景

• 想为科研添砖加瓦，因此需要一个好看的可视化。
• 或者只是想了解一下模型的效果。

• 希望快一点。

  • 因为使用markdown作为展示图片的方式，因此无需画图，速度大幅提高。

  • 我之前写了一个，一秒钟大概合成几张图，慢死，我自己都不用。

  • 现在这个只用九秒钟，3543x3384的大小。

  • 此外，同时提供各图像对应的ID信息。

使用示例

• 开源地址：asdTools/asdTools/Tools/Image/VisualizeRankOfReID.py at main
• 处理结果示例：asdTools/Sample/VisualizeRankOfReID at main
• 在test代码的evaluate()函数前，实例化一个VisualizeRankOfReID对象，将 距离矩阵、queryloader、galleryloader 传入即可。
  • 或者说，计算CMC和mAP前。
  • 使用示例见if __name__ == "__main__":

from asdTools.Classes.Tool.MarkdownTable import MarkdownTable
from asdTools.Classes.Base.BaseModel import BaseModel
import numpy as np

class VisualizeRankOfReID(BaseModel):
    """ Sample: Sample/VisualizeRankOfReID
    使用markdown表格进行ReID的Rank可视化。速度快，我的笔记本运行一次3543x3384(query-gallery)的可视化耗时12秒。
    Visualize ReID Rank by markdown tables. Fast - It takes only 12 seconds to generate visualizations on my laptop for a 3543x3384 query-gallery."
    """
    def __init__(self, **kwargs) -> None:
        super().__init__(multipleFiles=True, **kwargs)

    def __call__(self, 
            distmat, 
            query_loader, 
            gallery_loader, 
            top_k:int=10,
            pid_index:int=1,
            camid_index:int=2) -> str:
        self.run(distmat, query_loader, gallery_loader, top_k, pid_index, camid_index)

    def run(self, 
            distmat, 
            query_loader, 
            gallery_loader, 
            top_k,
            pid_index,
            camid_index) -> str:
        self.begining(isSimple=True)
        if top_k == -1: 
            self.warning("top_k is set to -1, all images will be shown, may too large.")
            top_k = distmat.shape[0]
        index2path = "index, good ratio, pid, camid, "
        for i in range(len(gallery_loader.dataset.dataset[0])-3):
            index2path += f"info[{i+3}], "
        index2path += "path \n"
        # ---------- Init Guery Datainfo ----------
        g_imgs = []
        g_pids = np.array([])
        g_camids = np.array([])
        for i, data_info in enumerate(gallery_loader.dataset.dataset):
            g_imgs.append(data_info[0])
            g_pids = np.append(g_pids, data_info[pid_index])
            g_camids = np.append(g_camids, data_info[camid_index])
        # ---------- Start Visualazation ----------
        for i, data_info in enumerate(query_loader.dataset.dataset):
            # ---------- Init Query Datainfo ----------
            q_img_path = data_info[0]
            q_pid = data_info[pid_index]
            q_camid = data_info[camid_index]
            # ---------- Sort Rank ----------
            distance_indices = np.argsort(distmat[i])
            query_index = np.argwhere(g_pids==q_pid)
            camera_index = np.argwhere(g_camids==q_camid)
            junk_index = np.intersect1d(query_index, camera_index)
            mask = np.in1d(distance_indices, junk_index, invert=True)
            sorted_indices = np.argwhere(mask==True).flatten()
            # ---------- Save Rank ----------
            # ----- rank gallery of i_query
            # rank gallery imgs
            g_imgs_path = [gallery_loader.dataset.dataset[distance_indices[j]][0] for j in sorted_indices[:top_k]]
            # log imgs
            md_table = MarkdownTable(["i", "query"] + [f"top_{j+1}" for j in range(top_k)])
            row_imgs = [i + 1]
            row_imgs.append(md_table.convert_imgPath_to_MDImgPath(self.convert_path_to_abspath(q_img_path)))
            row_imgs.extend([md_table.convert_imgPath_to_MDImgPath(self.convert_path_to_abspath(g_imgs_path[j])) for j in range(top_k)])
            md_table.add_row(row_imgs)
            # log id
            for j in range(len(data_info)):
                if j == 0: continue
                elif j == 1: row_ID = ["pid"]
                elif j == 2: row_ID = ["camid"]
                else: row_ID = [f"info[{j}]"]
                row_ID.append(data_info[j])
                row_ID.extend([gallery_loader.dataset.dataset[distance_indices[k]][j] for k in sorted_indices[:top_k]])
                md_table.add_row(row_ID)
            # log good ratio
            good_num = 0
            for j in range(top_k):
                idx = distance_indices[j]
                if g_pids[idx] == q_pid:
                    good_num += 1
            good_ratio = int(good_num / top_k * 100)
            # ----- map i_query to img_path of i_query
            index2path += f"{i+1}, {good_ratio}%, {q_pid}, {q_camid}, "
            for j in range(len(data_info)-3):
                index2path += f"{data_info[j+3]}, "
            index2path += f"{self.convert_path_to_abspath(q_img_path)} \n"
            # save markdown table
            md_table_path = self.generate_output_path(output_middle_dir=self._time_start, output_file=f"{i+1}-rank_table-{good_ratio}%good.md")
            self.save_file(md_table.output(), md_table_path)
        # ---------- Save Mapping ----------
        index2path_path = self.generate_output_path(output_middle_dir=self._time_start, output_file=f"0_index2path.csv")
        self.save_file(index2path, index2path_path)
        self.log(f"For find image faster, see mapping in {index2path_path}")
        self.done(isSimple=True)

if __name__ == "__main__":
    # All you need is place the codes before evaluate(distmat, ...) in test()

    # ---------- Sample 1 ----------
    """ When specify log_dir

def test_prcc(model, queryloader_same, queryloader_diff, galleryloader, dataset):
    logger = logging.getLogger('reid.test')
    ...

    from asdTools.Tools.Image.VisualizeRankOfReID import VisualizeRankOfReID
    rank_vis = VisualizeRankOfReID()
    log_dir = rank_vis.get_loggingLogger_path(logger)
    log_dir = rank_vis.join(log_dir, "Rank Visualization")
    VisualizeRankOfReID(log_dir=log_dir)(distmat, queryloader, galleryloader)

    logger.info("Computing CMC and mAP for the same clothes setting")
    cmc, mAP = evaluate(distmat_same, qs_pids, g_pids, qs_camids, g_camids)
    """

    # ---------- Sample 2 ----------
    """ Just visualization

def test_prcc(model, queryloader_same, queryloader_diff, galleryloader, dataset):

    from asdTools.Tools.Image.VisualizeRankOfReID import VisualizeRankOfReID
    VisualizeRankOfReID()(distmat, queryloader, galleryloader)

    logger.info("Computing CMC and mAP for the same clothes setting")
    cmc, mAP = evaluate(distmat_same, qs_pids, g_pids, qs_camids, g_camids)
    """