Human3.6m数据集预处理的一个小步骤

https://github.com/karfly/learnable-triangulation-pytorch/blob/master/mvn/datasets/human36m_preprocessing/README.md

根据论文提供的源码learnable-triangulation-pytorch

After that, you should have images unpacked as e.g. $THIS_REPOSITORY/data/human36m/processed/S1/Phoning-1/imageSequence/54138969/img_000001.jpg.

步骤二需要的是将图片按照Subject、camera、action这几个标签放到文件夹下。

而从网上下载的Human3.6m图片数据集是这样的，就是给出了7个Subject，并且用以下三个作为图片名称的代表。

S1_Directions.54138969_000001.jpg

S1_SittingDown_2.55011271_001511.jpg

S1_Directions_1.54138969_000001.jpg

subject+action+占位符+相机名称+帧数

为了实现论文Readme的要求撰写以下代码用于文件的分类复制

import os
import shutil
import numpy as np
import pandas as pd
#存放方式：S1...这些文件夹下有一个总的图片存放文件
#任务1：获取所有文件名，并将文件名拆分，并按动作和分隔符+相机名称进行分类，并存放到对应文件夹下
#例如S1_Directions_1.54138969_000001.jpg，将被存到 ..//S1//Directions_1//imageSequence//54138969

#任务2：对每个动作中的文件进行重命名img_+“编号”的形式
retval = {
    'subject_names': ['S1', 'S5', 'S6', 'S7', 'S8', 'S9', 'S11'],
    'camera_names': ['54138969', '55011271', '58860488', '60457274'],
    'action_names': [
        'Directions-1', 'Directions-2',
        'Discussion-1', 'Discussion-2',
        'Eating-1', 'Eating-2',
        'Greeting-1', 'Greeting-2',
        'Phoning-1', 'Phoning-2',
        'Posing-1', 'Posing-2',
        'Purchases-1', 'Purchases-2',
        'Sitting-1', 'Sitting-2',
        'SittingDown-1', 'SittingDown-2',
        'Smoking-1', 'Smoking-2',
        'TakingPhoto-1', 'TakingPhoto-2',
        'Waiting-1', 'Waiting-2',
        'Walking-1', 'Walking-2',
        'WalkingDog-1', 'WalkingDog-2',
        'WalkingTogether-1', 'WalkingTogether-2']
}
#S1_Directions.54138969_000001
#子集+动作+数据编号+图片
root_path = "E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed"
#起始地址
float_path = "E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed\\{}\\all".format(retval['subject_names'][6])
## 目标地址
DEST_path = "E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed\\{}\\{}\\imageSequence\\{}"
# float_path_1 = "E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed\\{}\\{}\\imageSequence\\{}".format(retval['action_names'][0], retval['camera_names'][0])

# float_path_2 = "E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed\\S1\\{}\\imageSequence\\{}".format(retval['action_names'][2*action_number+1])
# print(float_path_1)


#公共地址

new_name_prefix = "img_"

img_name = []

#把S1中的4个相机里面的图片挨个读一遍
# def path_read(i):
#     data_path = os.path.join(float_path, retval['camera_names'][i])
#     #很神奇，加上这句折后返回值直接变成文件了
#     # data_path = os.listdir(data_path)
#     return data_path
#
###############获取多个文件的文件名
# def get_image_filenames(folder_paths):
#     #放整个名称
#     df_name = pd.DataFrame(columns=['ImageName'])
#     #放细节
#     df_detail = pd.DataFrame(columns=['Subject', 'Action_', '_', 'Camera', 'FrameNumber'])
#     # 获取文件夹下的所有图片文件名
#     image_names = [name for name in os.listdir(folder_paths) if name.lower().endswith('.jpg')]
#     # 将数据添加到 DataFrame
#     data = {'ImageName': image_names}
#     df_name = df_name.append(pd.DataFrame(data), ignore_index=True)
#     for index, row in df_name.iterrows():
#         image_name = row['ImageName']
#         if image_name.endswith('.jpg'):
#             #计算下划线的个数区分是1还是0号数据集中的数据（3是1,2是0）
#             # 去除尾缀
#             image_name = image_name.rstrip('.jpg')
#             underscore_count = image_name.count('_')
#             if underscore_count == 2:
#                 #此时占位符为0
#                 _ = '0'
#                 # 按点进行分割
#                 segments = image_name.split('.')
#                 # 提取前两个片段
#                 subject_action, camera_frame = segments[:2]
#                 # 再按下划线进行分割，提取相机和帧编号
#                 subject, action = subject_action.split('_')
#                 # 再按点进行分割，提取相机和帧编号
#                 camera, frame_number = camera_frame.split('_')
#                 Action_ = '-'.join([action, _])
#                 # 将结果添加到 DataFrame
#                 df_detail = df_detail.append({'Subject': subject, 'Action_': Action_, '_': _, 'Camera': camera, 'FrameNumber': frame_number},
#                                ignore_index=True)
#             else:
#                 #下划线和点来分割文件名
#                 segments = image_name.split('_')
#                 # 提取前四个片段
#                 subject, action, camera_, frame_number = segments[:4]
#                 _, camera = camera_.split('.')
#                 Action_ = '-'.join([action, _])
#                 # 将结果添加到 DataFrame
#                 df_detail = df_detail.append({'Subject': subject, 'Action_': Action_, '_': _, 'Camera': camera, 'FrameNumber': frame_number},
#                                ignore_index=True)
#             Action, Camera = df_detail['Action_'], df_detail['Camera']
#     return Action, Camera, _

####获取单个文件的文件名
def get_image_filename(folder_paths):
    #basename就是获取'E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed\\S1\\all\\11\\S1_Directions.54138969_000001.jpg'中的S1_Directions.54138969_000001.jpg
    image_name = os.path.basename(folder_paths)
    # 将数据添加到 DataFrame
    #计算下划线的个数区分是1还是0号数据集中的数据（3是1,2是0）
    # 去除尾缀
    image_name = image_name.rstrip('.jpg')
    underscore_count = image_name.count('_')
    if underscore_count == 2:
        #此时占位符为0
        _ = '0'
        # 按点进行分割
        segments = image_name.split('.')
        # 提取前两个片段
        subject_action, camera_frame = segments[:2]
        # 再按下划线进行分割，提取相机和帧编号
        subject, action = subject_action.split('_')
        # 再按点进行分割，提取相机和帧编号
        camera, frame_number = camera_frame.split('_')
        Action_ = '-'.join([action, _])
        # 将结果添加到 DataFrame
    else:
        #下划线和点来分割文件名
        segments = image_name.split('_')
        # 提取前四个片段
        subject, action, camera_, frame_number = segments[:4]
        _, camera = camera_.split('.')
        Action_ = '-'.join([action, _])
        # 将结果添加到 DataFrame

    return Action_, camera, _, subject

#单个动作的移动
# def copy_images(source_root, dest_root, camera_folder,action_name):
def copy_images(source_root, dest_root,img_names, num):
    #当源路径是多个相机的根时候这样搞，还不行~
    # source_folder = [os.path.join(source_root, folder) for folder in camera_folder]
    # dest_folder = [os.path.join(dest_root, folder) for folder in camera_folder]
    # source_folder = source_root
    # dest_folder = dest_root
    # img_name_S1, img_detail_S1 = get_image_filenames(source_folder)
    # image_files = [f for f in os.listdir(source_folder) if f.endswith('.jpg')]
    # # # 创建目标文件夹
    # # os.makedirs(dest_folder, exist_ok=True)
    #
    # # # 获取源文件夹下所有的图片文件名
    # # image_files = [f for f in os.listdir(source_folder) if f.endswith('.jpg')]
    #
    # # 复制对应动作图片到目标文件夹

    # source_path = os.path.join(source_folder, image_file)
    # dest_path = os.path.join(dest_folder, image_file)
    for time in range(len(num)):
        source_path = os.path.join(source_root, img_names["ImageName"][num[time]])
        # dest_path = os.path.join(dest_root[0], img_names["ImageName"][num[time]])
        dest_path = os.path.join(dest_root, img_names["ImageName"][num[time]])
        shutil.copyfile(source_path, dest_path)

    # for index in img_detail:
    #     if img_detail['Action_'] == retval['action_names'][0]:
    #         i = img_detail.index[img_detail['Action_'] == retval['action_names'][0]].tolist()
    #         dest_path = DEST_path.format(retval['subject_names'][0], retval['action_names'][0],
    #                                      retval['camera_names'][0])
    #
    #         source_path = os.path.join(source_root, img_names[i])
    #         dest_path = os.path.join(dest_root[0], img_names[i])
    #         dest_paths.append(dest_path)
    #         source_paths.append(source_path)
    #         print(i)
    #         print(index)
    #         print(source_paths)
    #         print(dest_paths)
    # shutil.copyfile(source_root, dest_root)
# def rename(folder_path, prefix_length):
#     folder_path = os.path.join(base_folder, folder)






def classify_and_pack(folder_path, prefix_length):
    for filename in os.listdir(folder_path):
        if filename.endswith('.jpg'):
            # 获取前缀部分
            prefix = filename[:prefix_length]
            # 创建目标文件夹
            target_folder = os.path.join(folder_path, prefix)
            os.makedirs(target_folder, exist_ok=True)
            # 移动文件
            source_path = os.path.join(folder_path, filename)
            target_path = os.path.join(target_folder, filename)
            shutil.move(source_path, target_path)


# def rename_file(folder_path, prefix_length, suffix_length):
#     for filename in os.listdir(folder_path):
#         #获取文件命中的编号部分
if __name__ == '__main__':
    dest_paths = []
    source_paths = []
    inx = 0
    # camera_folder = np.array([path_read(0), path_read(1), path_read(2), path_read(3)])
    # img_name_S1, img_detail_S1 = get_image_filenames(camera_folder)
    # img_name_S1, img_detail_S1 = get_image_filenames(float_path)
    # img_name, img_detail = get_image_filenames(float_path)
    #提取动作和相机标签
    for filename in os.listdir(float_path):
        source_path = os.path.join(float_path, filename)
        # Action, Camera, _ = get_image_filenames(float_path)
        if filename.endswith('.jpg'):
            Action, Camera, _, Subject = get_image_filename(source_path)
            #知识点1，itertuples()这个是dataframe的，要让它变成在in之后的循环常量
            #知识点2，zip能让series类型变成in之后的循环常量
            #知识点3，zip（str1，str2），其中str1和str2是两个字符串数组，需要保证两者的元素个数一样才能打包
            if _ != '0':
                #知识点3，"E:\\project\\learnable-triangulation-pytorch-master\\data\\human36m\\processed\\{}\\{}\\imageSequence\\{}".format(1,2,3)可以实现地址的补全
                #source_path = os.path.join(folder_path, filename)这也是一种补全方法
                dest_folder = DEST_path.format(Subject, Action, Camera)
                #创建目标文件夹！可以一次创建多级目录
                os.makedirs(dest_folder, exist_ok=True)
                dest_path = os.path.join(dest_folder, filename)
                shutil.copyfile(source_path, dest_path)
        else:
            # 跳过不符合条件的文件
            print(f"Skipping: {filename} (not a .jpg file)")

    # for i, filename in len(img_name), os.listdir(float_path):
    #     if img_detail['Action_'][i] == retval['action_names'][0]:
    #         #代表满足当前动作的图片总数
    #         k = img_detail.index[img_detail['Action_'] == retval['action_names'][0]].tolist()
    #         if img_detail['Camera'][i] == retval['camera_names'][0]:
    #             #代表即满足当前图片又满足相机编号的图片总数
    #             j = img_detail.index[img_detail['Camera'] == retval['camera_names'][0]].tolist()
    #             dest_path = DEST_path.format(retval['subject_names'][0], retval['action_names'][0],
    #                                          img_detail['Camera'][i])
    #
    #             dest_paths.append(dest_path)
    #             inx = inx + 1
    #             copy_images(float_path, dest_paths[0], img_name, i)
    #         elif img_detail['Camera'][i] == retval['camera_names'][1]:
    #             print("良好")
    #         elif img_detail['Camera'][i] == retval['camera_names'][2]:
    #             print("cha")
    #         else:
    #             print("s")

    # copy_images(float_path, dest_paths)
    # dest_path = DEST_path.format(retval['subject_names'][0], retval['action_names'][0],
    #                              retval['camera_names'][0])
    # print(final_path)
    # copy_images(dest_paths[0], dest_root)


    # if img_detail['Action_'].astype(str) == retval['action_names'][0]:
    #     dest_path = DEST_path.format(retval['subject_names'][0],retval['action_names'][0], retval['camera_names'][0])

    # print(s)
    # for i in path_S1:
    #     temporary = get_image_filenames(i)
    #     img_name = img_name.append(temporary)
    # print(img_name_S1)