简介
介绍如何使用CPM（Convolutional Pose Machines）实现服饰关键点定位
原理
关键点定位是一类常见而有用的任务，某种意义上可以理解为一种特征工程
- 人脸关键点定位，可用于人脸识别、表情识别
- 人体骨骼关键点定位，可用于姿态估计
- 手部关键点定位，可用于手势识别
输入是一张图片，输出是每个关键点的x、y坐标，一般会归一化到0～1区间中，所以可以理解为回归问题
但是直接对坐标值进行回归会导致较大误差，更好的做法是输出一个低分辨率的热图，使得关键点所在位置输出较高响应，而其他位置则输出较低响应
CPM（Convolutional Pose Machines）的基本思想是使用多个级联的stage，每个stage包含多个CNN并且都输出热图
通过最小化每个stage的热图和ground truth之间的差距，从而得到越来越准确的关键点定位结果
数据
使用天池FashionAI全球挑战赛提供的数据，
其中服饰关键点定位赛题提供的训练集包括4W多张图片，测试集包括将近1W张图片
每张图片都指定了对应的服饰类别，共5类：上衣（blouse）、外套（outwear）、连身裙（dress）、半身裙（skirt）、裤子（trousers）
训练集还提供了每张图片对应的24个关键点的标注，包括x坐标、y坐标、是否可见三项信息，但并不是每类服饰都有24个关键点
关于以上数据的更多介绍可以参考以下文章，
为了简化问题，以下仅使用dress类别的训练集数据训练CPM模型
实现
加载库
# -*- coding: utf-8 -*-import tensorflow as tfimport numpy as npimport pandas as pdfrom sklearn.utils import shufflefrom sklearn.model_selection import train_test_splitimport cv2import matplotlib.pyplot as plt%matplotlib inlinefrom imageio import imread, imsaveimport osimport globfrom tqdm import tqdmimport warningswarnings.filterwarnings('ignore')复制代码 加载训练集和测试集
train = pd.read_csv(os.path.join('data', 'train', 'train.csv'))train_warm = pd.read_csv(os.path.join('data', 'train_warm', 'train_warm.csv'))test = pd.read_csv(os.path.join('data', 'test', 'test.csv'))print(len(train), len(train_warm), len(test))columns = train.columnsprint(len(columns), columns)train['image_id'] = train['image_id'].apply(lambda x:os.path.join('train', x))train_warm['image_id'] = train_warm['image_id'].apply(lambda x:os.path.join('train_warm', x))train = pd.concat([train, train_warm])del train_warmtrain.head()复制代码 仅保留dress类别的数据
train = train[train.image_category == 'dress']test = test[test.image_category == 'dress']print(len(train), len(test))复制代码
拆分标注信息中的x坐标、y坐标、是否可见
for col in columns: if col in ['image_id', 'image_category']: continue train[col + '_x'] = train[col].apply(lambda x:float(x.split('_')[0])) train[col + '_y'] = train[col].apply(lambda x:float(x.split('_')[1])) train[col + '_s'] = train[col].apply(lambda x:float(x.split('_')[2])) train.drop([col], axis=1, inplace=True)train.head()复制代码 将x坐标和y坐标进行归一化
features = [ 'neckline_left', 'neckline_right', 'center_front', 'shoulder_left', 'shoulder_right', 'armpit_left', 'armpit_right', 'waistline_left', 'waistline_right', 'cuff_left_in', 'cuff_left_out', 'cuff_right_in', 'cuff_right_out', 'hemline_left', 'hemline_right']train = train.to_dict('records')for i in tqdm(range(len(train))): record = train[i] img = imread(os.path.join('data', record['image_id'])) h = img.shape[0] w = img.shape[1] for col in features: if record[col + '_s'] >= 0: train[i][col + '_x'] /= w train[i][col + '_y'] /= h else: train[i][col + '_x'] = 0 train[i][col + '_y'] = 0复制代码 随机选一些训练数据并绘图查看
img_size = 256r = 10c = 10puzzle = np.ones((img_size * r, img_size * c, 3))random_indexs = np.random.choice(len(train), 100)for i in range(100): record = train[random_indexs[i]] img = imread(os.path.join('data', record['image_id'])) img = cv2.resize(img, (img_size, img_size)) for col in features: if record[col + '_s'] >= 0: cv2.circle(img, (int(img_size * record[col + '_x']), int(img_size * record[col + '_y'])), 3, (120, 240, 120), 2) img = img / 255. r = i // 10 c = i % 10 puzzle[r * img_size: (r + 1) * img_size, c * img_size: (c + 1) * img_size, :] = imgplt.figure(figsize=(15, 15))plt.imshow(puzzle)plt.show()复制代码 整理数据并分割训练集和验证集
X_train = []Y_train = []for i in tqdm(range(len(train))): record = train[i] img = imread(os.path.join('data', record['image_id'])) img = cv2.resize(img, (img_size, img_size)) y = [] for col in features: y.append([record[col + '_x'], record[col + '_y']]) X_train.append(img) Y_train.append(y)X_train = np.array(X_train)Y_train = np.array(Y_train)print(X_train.shape)print(Y_train.shape)X_train, X_valid, Y_train, Y_valid = train_test_split(X_train, Y_train, test_size=0.1)print(X_train.shape, Y_train.shape)print(X_valid.shape, Y_valid.shape)复制代码 定义一些参数和模型输入
batch_size = 16heatmap_size = 32stages = 6y_dim = Y_train.shape[1]X = tf.placeholder(tf.float32, [None, img_size, img_size, 3], name='X')Y = tf.placeholder(tf.float32, [None, heatmap_size, heatmap_size, y_dim + 1], name='Y')def conv2d(inputs, filters, kernel_size, padding='same', activation=tf.nn.relu, name=''): if name: return tf.layers.conv2d(inputs, filters=filters, kernel_size=kernel_size, strides=1, padding=padding, activation=activation, name=name, kernel_initializer=tf.contrib.layers.xavier_initializer()) else: return tf.layers.conv2d(inputs, filters=filters, kernel_size=kernel_size, strides=1, padding=padding, activation=activation, kernel_initializer=tf.contrib.layers.xavier_initializer())def maxpool2d(inputs): return tf.layers.max_pooling2d(inputs, pool_size=2, strides=2, padding='valid')复制代码
定义CPM模型，使用6个stage
stage_heatmaps = []# sub_stageh0 = maxpool2d(conv2d(conv2d(X, 64, 3), 64, 3))h0 = maxpool2d(conv2d(conv2d(h0, 128, 3), 128, 3))h0 = maxpool2d(conv2d(conv2d(conv2d(conv2d(h0, 256, 3), 256, 3), 256, 3), 256, 3))for i in range(6): h0 = conv2d(h0, 512, 3)sub_stage = conv2d(h0, 128, 3) # batch_size, 32, 32, 128# stage_1h0 = conv2d(sub_stage, 512, 1, padding='valid')h0 = conv2d(h0, y_dim + 1, 1, padding='valid', activation=None, name='stage_1')stage_heatmaps.append(h0)# other stagesfor stage in range(2, stages + 1): h0 = tf.concat([stage_heatmaps[-1], sub_stage], axis=3) for i in range(5): h0 = conv2d(h0, 128, 7) h0 = conv2d(h0, 128, 1, padding='valid') h0 = conv2d(h0, y_dim + 1, 1, padding='valid', activation=None, name='stage_%d' % stage) stage_heatmaps.append(h0)复制代码
定义损失函数和优化器
global_step = tf.Variable(0, trainable=False)learning_rate = tf.train.exponential_decay(0.001, global_step=global_step, decay_steps=1000, decay_rate=0.9)losses = [0 for _ in range(stages)]total_loss = 0for stage in range(stages): losses[stage] = tf.losses.mean_squared_error(Y, stage_heatmaps[stage]) total_loss += losses[stage]total_loss_with_reg = total_loss + tf.contrib.layers.apply_regularization(tf.contrib.layers.l2_regularizer(1e-10), tf.trainable_variables())total_loss = total_loss / stagestotal_loss_with_reg = total_loss_with_reg / stagesoptimizer = tf.contrib.layers.optimize_loss(total_loss_with_reg, global_step=global_step, learning_rate=learning_rate, optimizer='Adam', increment_global_step=True)复制代码
由于训练数据较少，因此定义一个数据增强的函数
def transform(X_batch, Y_batch): X_data = [] Y_data = [] offset = 20 for i in range(X_batch.shape[0]): img = X_batch[i] # random rotation degree = int(np.random.random() * offset - offset / 2) rad = degree / 180 * np.pi mat = cv2.getRotationMatrix2D((img_size / 2, img_size / 2), degree, 1) img_ = cv2.warpAffine(img, mat, (img_size, img_size), borderValue=(255, 255, 255)) # random translation x0 = int(np.random.random() * offset - offset / 2) y0 = int(np.random.random() * offset - offset / 2) mat = np.float32([[1, 0, x0], [0, 1, y0]]) img_ = cv2.warpAffine(img_, mat, (img_size, img_size), borderValue=(255, 255, 255)) # random flip if np.random.random() > 0.5: img_ = np.fliplr(img_) flip = True else: flip = False X_data.append(img_) points = [] for j in range(y_dim): x = Y_batch[i, j, 0] * img_size y = Y_batch[i, j, 1] * img_size # random rotation dx = x - img_size / 2 dy = y - img_size / 2 x = int(dx * np.cos(rad) + dy * np.sin(rad) + img_size / 2) y = int(-dx * np.sin(rad) + dy * np.cos(rad) + img_size / 2) # random translation x += x0 y += y0 x = x / img_size y = y / img_size points.append([x, y]) # random flip if flip: data = {features[j]: points[j] for j in range(y_dim)} points = [] for j in range(y_dim): col = features[j] if col.find('left') >= 0: col = col.replace('left', 'right') elif col.find('right') >= 0: col = col.replace('right', 'left') [x, y] = data[col] x = 1 - x points.append([x, y]) Y_data.append(points) X_data = np.array(X_data) Y_data = np.array(Y_data) # preprocess X_data = (X_data / 255. - 0.5) * 2 Y_heatmap = [] for i in range(Y_data.shape[0]): heatmaps = [] invert_heatmap = np.ones((heatmap_size, heatmap_size)) for j in range(Y_data.shape[1]): x0 = int(Y_data[i, j, 0] * heatmap_size) y0 = int(Y_data[i, j, 1] * heatmap_size) x = np.arange(0, heatmap_size, 1, float) y = x[:, np.newaxis] cur_heatmap = np.exp(-((x - x0) ** 2 + (y - y0) ** 2) / (2.0 * 1.0 ** 2)) heatmaps.append(cur_heatmap) invert_heatmap -= cur_heatmap heatmaps.append(invert_heatmap) Y_heatmap.append(heatmaps) Y_heatmap = np.array(Y_heatmap) Y_heatmap = np.transpose(Y_heatmap, (0, 2, 3, 1)) # batch_size, heatmap_size, heatmap_size, y_dim + 1 return X_data, Y_data, Y_heatmap复制代码 查看数据增强之后的图片和关键点是否正确
X_batch = X_train[:batch_size]Y_batch = Y_train[:batch_size]X_data, Y_data, Y_heatmap = transform(X_batch, Y_batch)n = int(np.sqrt(batch_size))puzzle = np.ones((img_size * n, img_size * n, 3))for i in range(batch_size): img = (X_data[i] + 1) / 2 for j in range(y_dim): cv2.circle(img, (int(img_size * Y_data[i, j, 0]), int(img_size * Y_data[i, j, 1])), 3, (120, 240, 120), 2) r = i // n c = i % n puzzle[r * img_size: (r + 1) * img_size, c * img_size: (c + 1) * img_size, :] = imgplt.figure(figsize=(12, 12))plt.imshow(puzzle)plt.show()复制代码
可以看到，在经过随机旋转、随机平移、随机水平翻转之后，图片和关键点依旧一一对应
训练模型，使用early stopping
sess = tf.Session()sess.run(tf.global_variables_initializer())OUTPUT_DIR = 'samples'if not os.path.exists(OUTPUT_DIR): os.mkdir(OUTPUT_DIR)for stage in range(stages): tf.summary.scalar('loss/loss_stage_%d' % (stage + 1), losses[stage])tf.summary.scalar('loss/total_loss', total_loss)summary = tf.summary.merge_all()writer = tf.summary.FileWriter(OUTPUT_DIR)loss_valid_min = np.infsaver = tf.train.Saver()epochs = 100patience = 10for e in range(epochs): loss_train = [] loss_valid = [] X_train, Y_train = shuffle(X_train, Y_train) for i in tqdm(range(X_train.shape[0] // batch_size)): X_batch = X_train[i * batch_size: (i + 1) * batch_size, :, :, :] Y_batch = Y_train[i * batch_size: (i + 1) * batch_size, :, :] X_data, Y_data, Y_heatmap = transform(X_batch, Y_batch) _, ls, lr, stage_heatmaps_ = sess.run([optimizer, total_loss, learning_rate, stage_heatmaps], feed_dict={X: X_data, Y: Y_heatmap}) loss_train.append(ls) if i > 0 and i % 100 == 0: writer.add_summary(sess.run(summary, feed_dict={X: X_data, Y: Y_heatmap}), e * X_train.shape[0] // batch_size + i) writer.flush() loss_train = np.mean(loss_train) demo_img = (X_data[0] + 1) / 2 demo_heatmaps = [] for stage in range(stages): demo_heatmap = stage_heatmaps_[stage][0, :, :, :y_dim].reshape((heatmap_size, heatmap_size, y_dim)) demo_heatmap = cv2.resize(demo_heatmap, (img_size, img_size)) demo_heatmap = np.amax(demo_heatmap, axis=2) demo_heatmap = np.reshape(demo_heatmap, (img_size, img_size, 1)) demo_heatmap = np.repeat(demo_heatmap, 3, axis=2) demo_heatmaps.append(demo_heatmap) demo_gt_heatmap = Y_heatmap[0, :, :, :y_dim].reshape((heatmap_size, heatmap_size, y_dim)) demo_gt_heatmap = cv2.resize(demo_gt_heatmap, (img_size, img_size)) demo_gt_heatmap = np.amax(demo_gt_heatmap, axis=2) demo_gt_heatmap = np.reshape(demo_gt_heatmap, (img_size, img_size, 1)) demo_gt_heatmap = np.repeat(demo_gt_heatmap, 3, axis=2) upper_img = np.concatenate((demo_heatmaps[0], demo_heatmaps[1], demo_heatmaps[2]), axis=1) blend_img = 0.5 * demo_img + 0.5 * demo_gt_heatmap lower_img = np.concatenate((demo_heatmaps[-1], demo_gt_heatmap, blend_img), axis=1) demo_img = np.concatenate((upper_img, lower_img), axis=0) imsave(os.path.join(OUTPUT_DIR, 'sample_%d.jpg' % e), demo_img) X_valid, Y_valid = shuffle(X_valid, Y_valid) for i in range(X_valid.shape[0] // batch_size): X_batch = X_valid[i * batch_size: (i + 1) * batch_size, :, :, :] Y_batch = Y_valid[i * batch_size: (i + 1) * batch_size, :, :] X_data, Y_data, Y_heatmap = transform(X_batch, Y_batch) ls = sess.run(total_loss, feed_dict={X: X_data, Y: Y_heatmap}) loss_valid.append(ls) loss_valid = np.mean(loss_valid) print('Epoch %d, lr %.6f, train loss %.6f, valid loss %.6f' % (e, lr, loss_train, loss_valid)) if loss_valid < loss_valid_min: print('Saving model...') saver.save(sess, os.path.join(OUTPUT_DIR, 'cpm')) loss_valid_min = loss_valid patience = 10 else: patience -= 1 if patience == 0: break复制代码 经过87个epoch之后训练提前停止，训练集损失为0.001228，验证集损失为0.001871，验证集最低损失为0.001821
训练集结果如下，上面三张图依次为stage1、stage2、stage3的结果，下面三张图依次为stage6、ground truth、groud truth和原图
在本机上使用以下代码，对测试集中的图片进行关键点定位
# -*- coding: utf-8 -*-import tensorflow as tfimport numpy as npimport pandas as pdfrom sklearn.utils import shuffleimport cv2import matplotlib.pyplot as pltfrom imageio import imread, imsaveimport ostest = pd.read_csv(os.path.join('data', 'test', 'test.csv'))test['image_id'] = test['image_id'].apply(lambda x:os.path.join('test', x))test = test[test.image_category == 'dress']test = test['image_id'].valuesbatch_size = 16img_size = 256test = shuffle(test)test = test[:batch_size]X_test = []for i in range(batch_size): img = imread(os.path.join('data', test[i])) img = cv2.resize(img, (img_size, img_size)) X_test.append(img)X_test = np.array(X_test)print(X_test.shape)sess = tf.Session()sess.run(tf.global_variables_initializer())OUTPUT_DIR = 'samples'saver = tf.train.import_meta_graph(os.path.join(OUTPUT_DIR, 'cpm.meta'))saver.restore(sess, tf.train.latest_checkpoint(OUTPUT_DIR))stages = 6y_dim = 15heatmap_size = 32graph = tf.get_default_graph()X = graph.get_tensor_by_name('X:0')stage_heatmap = graph.get_tensor_by_name('stage_%d/BiasAdd:0' % stages)def visualize_result(imgs, heatmap, joints): imgs = imgs.astype(np.int32) coords = [] for i in range(imgs.shape[0]): hp = heatmap[i, :, :, :joints].reshape((heatmap_size, heatmap_size, joints)) hp = cv2.resize(hp, (img_size, img_size)) coord = np.zeros((joints, 2)) for j in range(joints): xy = np.unravel_index(np.argmax(hp[:, :, j]), (img_size, img_size)) coord[j, :] = [xy[0], xy[1]] cv2.circle(imgs[i], (xy[1], xy[0]), 3, (120, 240, 120), 2) coords.append(coord) return imgs / 255., coordsheatmap = sess.run(stage_heatmap, feed_dict={X: (X_test / 255. - 0.5) * 2})X_test, coords = visualize_result(X_test, heatmap, y_dim) n = int(np.sqrt(batch_size))puzzle = np.ones((img_size * n, img_size * n, 3))for i in range(batch_size): img = X_test[i] r = i // n c = i % n puzzle[r * img_size: (r + 1) * img_size, c * img_size: (c + 1) * img_size, :] = imgplt.figure(figsize=(12, 12))plt.imshow(puzzle)plt.show()imsave('服饰关键点定位测试集结果.jpg', puzzle)复制代码 测试集结果如下，对于大多数情况能够得到很准确的关键点定位结果
参考
- Convolutional Pose Machines：
- Code repository for Convolutional Pose Machines：
- 天池FashionAI全球挑战赛小小尝试：