diff --git a/README.md b/README.md index beda92c..9597162 100644 --- a/README.md +++ b/README.md @@ -7,7 +7,7 @@ ### 1. 📣 数据介绍 -确定了业务场景之后，需要手机大量的数据（之前参加过一个安全帽识别检测的比赛，但是数据在比赛平台无法下载为己用），一般来说包含两大来源，一部分是网络数据，可以通过百度、Google图片爬虫拿到，另一部分是用户场景的视频录像，后一部分相对来说数据量更大，但出于商业因素几乎不会开放。本文开源的安全帽检测数据集([SafetyHelmetWearing-Dataset, SHWD](https://github.com/njvisionpower/Safety-Helmet-Wearing-Dataset))主要通过爬虫拿到，总共有7581张图像，包含9044个佩戴安全帽的bounding box（正类），以及111514个未佩戴安全帽的bounding box(负类)，所有的图像用labelimg标注出目标区域及类别。其中每个bounding box的标签：hat”表示佩戴安全帽，“person”表示普通未佩戴的行人头部区域的bounding box。另外本数据集中person标签的数据大多数来源于[SCUT-HEAD](https://github.com/HCIILAB/SCUT-HEAD-Dataset-Release)数据集，用于判断是未佩戴安全帽的人。大致说一下数据集构造的过程： +确定了业务场景之后，需要收集大量的数据（之前参加过一个安全帽识别检测的比赛，但是数据在比赛平台无法下载为己用），一般来说包含两大来源，一部分是网络数据，可以通过百度、Google图片爬虫拿到，另一部分是用户场景的视频录像，后一部分相对来说数据量更大，但出于商业因素几乎不会开放。本文开源的安全帽检测数据集([SafetyHelmetWearing-Dataset, SHWD](https://github.com/njvisionpower/Safety-Helmet-Wearing-Dataset))主要通过爬虫拿到，总共有7581张图像，包含9044个佩戴安全帽的bounding box（正类），以及111514个未佩戴安全帽的bounding box(负类)，所有的图像用labelimg标注出目标区域及类别。其中每个bounding box的标签：“hat”表示佩戴安全帽，“person”表示普通未佩戴的行人头部区域的bounding box。另外本数据集中person标签的数据大多数来源于[SCUT-HEAD](https://github.com/HCIILAB/SCUT-HEAD-Dataset-Release)数据集，用于判断是未佩戴安全帽的人。大致说一下数据集构造的过程： 1.数据爬取 @@ -47,7 +47,7 @@ Packages: - opencv-python - tqdm -将预训练的darknet的权重下载，下载地址：,并将该weight文件拷贝大`./data/darknet_weights/`下，因为这是darknet版本的预训练权重，需要转化为Tensorflow可用的版本，运行如下代码可以实现： +将预训练的darknet的权重下载，下载地址：,并将该weight文件拷贝到`./data/darknet_weights/`下，因为这是darknet版本的预训练权重，需要转化为Tensorflow可用的版本，运行如下代码可以实现： ```shell python convert_weight.py @@ -55,7 +55,7 @@ python convert_weight.py 这样转化后的Tensorflow checkpoint文件被存放在：`./data/darknet_weights/`目录。你也可以下载已经转化好的模型： -[Google云盘]((https://drive.google.com/drive/folders/1mXbNgNxyXPi7JNsnBaxEv1-nWr7SVoQt?usp=sharing) [GitHub Release](https://github.com/wizyoung/YOLOv3_TensorFlow/releases/) +[Google云盘](https://drive.google.com/drive/folders/1mXbNgNxyXPi7JNsnBaxEv1-nWr7SVoQt?usp=sharing) [GitHub Release](https://github.com/wizyoung/YOLOv3_TensorFlow/releases/) ### 3.🔰 训练数据构建 @@ -67,17 +67,19 @@ python convert_weight.py ```shell python data_pro.py ``` -分割训练集，验证集，测试集并在`./data/my_data/`下生成`train.txt/val.txt/test.txt`，对于一张图像对应一行数据，包括`image_index`,`image_absolute_path`,`box_1`,`box_2`,...,`box_n`,每个字段中间是用空格分隔的，其中: +分割训练集，验证集，测试集并在`./data/my_data/`下生成`train.txt/val.txt/test.txt`，对于一张图像对应一行数据，包括`image_index`,`image_absolute_path`, `img_width`, `img_height`,`box_1`,`box_2`,...,`box_n`,每个字段中间是用空格分隔的，其中: + `image_index`文本的行号 ++ `image_absolute_path` 一定是绝对路径 ++ `img_width`, `img_height`,`box_1`,`box_2`,...,`box_n`中涉及数值的取值一定取int型 + `box_x`的形式为：`label_index, x_min,y_min,x_max,y_max`(注意坐标原点在图像的左上角) + `label_index`是label对应的index(取值为0-class_num-1),这里要注意YOLO系列的模型训练与SSD不同，label不包含background 例子： ``` -0 xxx/xxx/a.jpg 0 453 369 473 391 1 588 245 608 268 -1 xxx/xxx/b.jpg 1 466 403 485 422 2 793 300 809 320 +0 xxx/xxx/a.jpg 1920,1080,0 453 369 473 391 1 588 245 608 268 +1 xxx/xxx/b.jpg 1920,1080,1 466 403 485 422 2 793 300 809 320 ... ``` @@ -98,6 +100,8 @@ person python get_kmeans.py ``` +![](docs/kmeans.png) + 可以得到9个anchors和平均的IOU,把anchors保存在文本文件：`./data/yolo_anchors.txt`, **注意: Kmeans计算出的YOLO Anchors是在在调整大小的图像比例的，默认的调整大小方法是保持图像的纵横比。** @@ -112,8 +116,8 @@ python get_kmeans.py 修改arg.py 

 ### Some paths
-train_file = './data/my_data/train.txt'  # The path of the training txt file.
-val_file = './data/my_data/val.txt'  # The path of the validation txt file.
+train_file = './data/my_data/label/train.txt'  # The path of the training txt file.
+val_file = './data/my_data/label/val.txt'  # The path of the validation txt file.
 restore_path = './data/darknet_weights/yolov3.ckpt'  # The path of the weights to restore.
 save_dir = './checkpoint/'  # The directory of the weights to save.
 log_dir = './data/logs/'  # The directory to store the tensorboard log files.
@@ -121,11 +125,12 @@ progress_log_path = './data/progress.log'  # The path to record the training pro
 anchor_path = './data/yolo_anchors.txt'  # The path of the anchor txt file.
 class_name_path = './data/coco.names'  # The path of the class names.
 ### Training releated numbers
-batch_size = 2  # 需要调整为自己的类别数
+batch_size = 32  #6
 img_size = [416, 416]  # Images will be resized to `img_size` and fed to the network, size format: [width, height]
-total_epoches = 500  # 训练周期调整
-train_evaluation_step = 50  # Evaluate on the training batch after some steps.
-val_evaluation_epoch = 1  # Evaluate on the whole validation dataset after some steps. Set to None to evaluate every epoch.
+letterbox_resize = True  # Whether to use the letterbox resize, i.e., keep the original aspect ratio in the resized image.
+total_epoches = 500
+train_evaluation_step = 100  # Evaluate on the training batch after some steps.
+val_evaluation_epoch = 50  # Evaluate on the whole validation dataset after some epochs. Set to None to evaluate every epoch.
 save_epoch = 10  # Save the model after some epochs.
 batch_norm_decay = 0.99  # decay in bn ops
 weight_decay = 5e-4  # l2 weight decay
@@ -134,45 +139,52 @@ global_step = 0  # used when resuming training
 num_threads = 10  # Number of threads for image processing used in tf.data pipeline.
 prefetech_buffer = 5  # Prefetech_buffer used in tf.data pipeline.
 ### Learning rate and optimizer
-optimizer_name = 'adam'  # Chosen from [sgd, momentum, adam, rmsprop]
+optimizer_name = 'momentum'  # Chosen from [sgd, momentum, adam, rmsprop]
 save_optimizer = True  # Whether to save the optimizer parameters into the checkpoint file.
-learning_rate_init = 1e-3
-lr_type = 'exponential'  # Chosen from [fixed, exponential, cosine_decay, cosine_decay_restart, piecewise]
+learning_rate_init = 1e-4
+lr_type = 'piecewise'  # Chosen from [fixed, exponential, cosine_decay, cosine_decay_restart, piecewise]
 lr_decay_epoch = 5  # Epochs after which learning rate decays. Int or float. Used when chosen `exponential` and `cosine_decay_restart` lr_type.
 lr_decay_factor = 0.96  # The learning rate decay factor. Used when chosen `exponential` lr_type.
 lr_lower_bound = 1e-6  # The minimum learning rate.
-# piecewise params
-pw_boundaries = [60, 80]  # epoch based boundaries
-pw_values = [learning_rate_init, 3e-5, 1e-4]
+# only used in piecewise lr type
+pw_boundaries = [30, 50]  # epoch based boundaries
+pw_values = [learning_rate_init, 3e-5, 1e-5]
 ### Load and finetune
 # Choose the parts you want to restore the weights. List form.
-# Set to None to restore the whole model.
-restore_part = ['yolov3/darknet53_body']
+# restore_include: None, restore_exclude: None  => restore the whole model
+# restore_include: None, restore_exclude: scope  => restore the whole model except `scope`
+# restore_include: scope1, restore_exclude: scope2  => if scope1 contains scope2, restore scope1 and not restore scope2 (scope1 - scope2)
+# choise 1: only restore the darknet body
+# restore_include = ['yolov3/darknet53_body']
+# restore_exclude = None
+# choise 2: restore all layers except the last 3 conv2d layers in 3 scale
+restore_include = None
+restore_exclude = ['yolov3/yolov3_head/Conv_14', 'yolov3/yolov3_head/Conv_6', 'yolov3/yolov3_head/Conv_22']
 # Choose the parts you want to finetune. List form.
 # Set to None to train the whole model.
 update_part = ['yolov3/yolov3_head']
 ### other training strategies
-multi_scale_train = False  # Whether to apply multi-scale training strategy. Image size varies from [320, 320] to [640, 640] by default.
-use_label_smooth = False # Whether to use class label smoothing strategy.
-use_focal_loss = False  # Whether to apply focal loss on the conf loss.
-use_mix_up = False  # Whether to use mix up data augmentation strategy. # 数据增强
+multi_scale_train = True  # Whether to apply multi-scale training strategy. Image size varies from [320, 320] to [640, 640] by default.
+use_label_smooth = True # Whether to use class label smoothing strategy.
+use_focal_loss = True  # Whether to apply focal loss on the conf loss.
+use_mix_up = True  # Whether to use mix up data augmentation strategy. 
 use_warm_up = True  # whether to use warm up strategy to prevent from gradient exploding.
 warm_up_epoch = 3  # Warm up training epoches. Set to a larger value if gradient explodes.
 ### some constants in validation
-# nms 非极大值抑制
-nms_threshold = 0.5  # iou threshold in nms operation
-score_threshold = 0.5  # threshold of the probability of the classes in nms operation
-nms_topk = 50  # keep at most nms_topk outputs after nms
+# nms
+nms_threshold = 0.45  # iou threshold in nms operation
+score_threshold = 0.01  # threshold of the probability of the classes in nms operation, i.e. score = pred_confs * pred_probs. set lower for higher recall.
+nms_topk = 150  # keep at most nms_topk outputs after nms
 # mAP eval
 eval_threshold = 0.5  # the iou threshold applied in mAP evaluation
+use_voc_07_metric = False  # whether to use voc 2007 evaluation metric, i.e. the 11-point metric
 ### parse some params
 anchors = parse_anchors(anchor_path)
 classes = read_class_names(class_name_path)
 class_num = len(classes)
 train_img_cnt = len(open(train_file, 'r').readlines())
 val_img_cnt = len(open(val_file, 'r').readlines())
-train_batch_num = int(math.ceil(float(train_img_cnt) / batch_size))  # iteration
-
+train_batch_num = int(math.ceil(float(train_img_cnt) / batch_size))
 lr_decay_freq = int(train_batch_num * lr_decay_epoch)
 pw_boundaries = [float(i) * train_batch_num + global_step for i in pw_boundaries]
diff --git a/args.py b/args.py index 7d67c82..51517f6 100644 --- a/args.py +++ b/args.py @@ -7,8 +7,8 @@ import math ### Some paths -train_file = './data/my_data/train.txt' # The path of the training txt file. -val_file = './data/my_data/val.txt' # The path of the validation txt file. +train_file = './data/my_data/label/train.txt' # The path of the training txt file. +val_file = './data/my_data/label/val.txt' # The path of the validation txt file. restore_path = './data/darknet_weights/yolov3.ckpt' # The path of the weights to restore. save_dir = './checkpoint/' # The directory of the weights to save. log_dir = './data/logs/' # The directory to store the tensorboard log files. diff --git a/data/coco.names b/data/coco.names index d32321c..3e8e951 100644 --- a/data/coco.names +++ b/data/coco.names @@ -1 +1,2 @@ -biopsy forceps \ No newline at end of file +hat +person \ No newline at end of file diff --git a/data/yolo_anchors.txt b/data/yolo_anchors.txt index 0281231..633e5d2 100644 --- a/data/yolo_anchors.txt +++ b/data/yolo_anchors.txt @@ -1 +1 @@ -676,197, 763,250, 684,283, 868,231, 745,273, 544,391, 829,258, 678,316, 713,355 \ No newline at end of file +5,5, 6,7, 7,9, 10,11, 13,15, 19,21, 27,31, 43,50, 79,93 \ No newline at end of file diff --git a/data_pro.py b/data_pro.py index 93ff560..e9562f2 100644 --- a/data_pro.py +++ b/data_pro.py @@ -33,12 +33,12 @@ def __init__(self,data_path): def load_labels(self, model): if model == 'train': - txtname = os.path.join(self.data_path, 'train_img.txt') + txtname = os.path.join(self.data_path, 'ImageSets/Main/train.txt') if model == 'test': - txtname = os.path.join(self.data_path, 'test_img.txt') + txtname = os.path.join(self.data_path, 'ImageSets/Main/test.txt') if model == "val": - txtname = os.path.join(self.data_path, 'val_img.txt') + txtname = os.path.join(self.data_path, 'ImageSets/Main/val.txt') with open(txtname, 'r') as f: @@ -47,14 +47,14 @@ def load_labels(self, model): my_index = 0 for ind in image_ind: - class_inds, x1s, y1s, x2s, y2s = self.load_data(ind) + class_inds, x1s, y1s, x2s, y2s,img_width,img_height = self.load_data(ind) if len(class_inds) == 0: pass else: annotation_label = "" #box_x: label_index, x_min,y_min,x_max,y_max - for label_i in range(len(clas_inds)): + for label_i in range(len(class_inds)): annotation_label += " " + str(class_inds[label_i]) annotation_label += " " + str(x1s[label_i]) @@ -62,8 +62,8 @@ def load_labels(self, model): annotation_label += " " + str(x2s[label_i]) annotation_label += " " + str(y2s[label_i]) - with open(model+".txt","a") as f: - f.write(str(my_index) + " " + data_path+"/ImageSets/"+ind+".jpg" + annotation_label + "\n") + with open("./data/my_data/label/"+model+".txt","a") as f: + f.write(str(my_index) + " " + data_path+"/JPEGImages/"+ind+".jpg"+" "+str(img_width) +" "+str(img_height)+ annotation_label + "\n") my_index += 1 @@ -76,8 +76,8 @@ def load_data(self, index): filename = os.path.join(self.data_path, 'Annotations', index + '.xml') tree = ET.parse(filename) image_size = tree.find('size') - # image_width = float(image_size.find('width').text) - # image_height = float(image_size.find('height').text) + image_width = int(float(image_size.find('width').text)) + image_height = int(float(image_size.find('height').text)) # h_ratio = 1.0 * self.image_size / image_height # w_ratio = 1.0 * self.image_size / image_width @@ -91,37 +91,38 @@ def load_data(self, index): for obj in objects: box = obj.find('bndbox') - x1 = float(box.find('xmin').text) - y1 = float(box.find('ymin').text) - x2 = float(box.find('xmax').text) - y2 = float(box.find('ymax').text) + x1 = int(float(box.find('xmin').text)) + y1 = int(float(box.find('ymin').text)) + x2 = int(float(box.find('xmax').text)) + y2 = int(float(box.find('ymax').text)) # x1 = max(min((float(box.find('xmin').text)) * w_ratio, self.image_size), 0) # y1 = max(min((float(box.find('ymin').text)) * h_ratio, self.image_size), 0) # x2 = max(min((float(box.find('xmax').text)) * w_ratio, self.image_size), 0) # y2 = max(min((float(box.find('ymax').text)) * h_ratio, self.image_size), 0) - class_ind = self.class_to_ind[obj.find('name').text] - # class_ind = self.class_to_ind[obj.find('name').text.lower().strip()] - - # boxes = [0.5 * (x1 + x2) / self.image_size, 0.5 * (y1 + y2) / self.image_size, np.sqrt((x2 - x1) / self.image_size), np.sqrt((y2 - y1) / self.image_size)] - # cx = 1.0 * boxes[0] * self.cell_size - # cy = 1.0 * boxes[1] * self.cell_size - # xind = int(np.floor(cx)) - # yind = int(np.floor(cy)) - - # label[yind, xind, :, 0] = 1 - # label[yind, xind, :, 1:5] = boxes - # label[yind, xind, :, 5 + class_ind] = 1 - - if x1 >= x2 or y1 >= y2: - pass - else: - class_inds.append(class_ind) - x1s.append(x1) - y1s.append(y1) - x2s.append(x2) - y2s.append(y2) - - return class_inds, x1s, y1s, x2s, y2s + if obj.find('name').text in self.classes: + class_ind = self.class_to_ind[obj.find('name').text] + # class_ind = self.class_to_ind[obj.find('name').text.lower().strip()] + + # boxes = [0.5 * (x1 + x2) / self.image_size, 0.5 * (y1 + y2) / self.image_size, np.sqrt((x2 - x1) / self.image_size), np.sqrt((y2 - y1) / self.image_size)] + # cx = 1.0 * boxes[0] * self.cell_size + # cy = 1.0 * boxes[1] * self.cell_size + # xind = int(np.floor(cx)) + # yind = int(np.floor(cy)) + + # label[yind, xind, :, 0] = 1 + # label[yind, xind, :, 1:5] = boxes + # label[yind, xind, :, 5 + class_ind] = 1 + + if x1 >= x2 or y1 >= y2: + pass + else: + class_inds.append(class_ind) + x1s.append(x1) + y1s.append(y1) + x2s.append(x2) + y2s.append(y2) + + return class_inds, x1s, y1s, x2s, y2s, image_width, image_height def data_split(img_path): @@ -141,19 +142,19 @@ def data_split(img_path): for file in files: if file in val_part: - with open("./data/my_data/val_img.txt","a") as val_f: + with open("./data/my_data/ImageSets/Main/val.txt","a") as val_f: val_f.write(file[:-4] + "\n" ) val_index += 1 elif file in test_part: - with open("./data/my_data/test_img.txt","a") as test_f: + with open("./data/my_data/ImageSets/Main/test.txt","a") as test_f: test_f.write(file[:-4] + "\n") test_index += 1 else: - with open("./data/my_data/train_img.txt","a") as train_f: + with open("./data/my_data/ImageSets/Main/train.txt","a") as train_f: train_f.write(file[:-4] + "\n") train_index += 1 @@ -166,12 +167,13 @@ def data_split(img_path): if __name__ == "__main__": # 分割train, val, test - img_path = "./data/my_data/ImageSets" - data_split(img_path) + # img_path = "./data/my_data/ImageSets" + # data_split(img_path) print("===========split data finish============") # 做YOLO V3需要的训练集 - data_path = "./data/my_data" # 尽量用绝对路径 + base_path = os.getcwd() + data_path = os.path.join(base_path,"data/my_data") # 绝对路径 data_p = Data_preprocess(data_path) data_p.load_labels("train") diff --git a/docs/kmeans.png b/docs/kmeans.png new file mode 100644 index 0000000..dcbc2c1 Binary files /dev/null and b/docs/kmeans.png differ diff --git a/get_kmeans.py b/get_kmeans.py index dc1faac..a8863f3 100644 --- a/get_kmeans.py +++ b/get_kmeans.py @@ -98,8 +98,8 @@ def parse_anno(annotation_path, target_size=None): result = [] for line in anno: s = line.strip().split(' ') - img_w = int(s[2]) - img_h = int(s[3]) + img_w = int(float(s[2])) + img_h = int(float(s[3])) s = s[4:] box_cnt = len(s) // 5 for i in range(box_cnt): @@ -139,7 +139,7 @@ def get_kmeans(anno, cluster_num=9): # if target_resize is speficied, the anchors are on the resized image scale # if target_resize is set to None, the anchors are on the original image scale target_size = [416, 416] - annotation_path = "./data/my_data/train.txt" + annotation_path = "./data/my_data/label/train.txt" anno_result = parse_anno(annotation_path, target_size=target_size) anchors, ave_iou = get_kmeans(anno_result, 9) diff --git a/utils/data_utils.py b/utils/data_utils.py index 3e22958..ba1cda6 100644 --- a/utils/data_utils.py +++ b/utils/data_utils.py @@ -128,6 +128,7 @@ def parse_data(line, class_num, img_size, anchors, mode, letterbox_resize): if not isinstance(line, list): img_idx, pic_path, boxes, labels, _, _ = parse_line(line) img = cv2.imread(pic_path) + # print(img.shape[:2]) # expand the 2nd dimension, mix up weight default to 1. boxes = np.concatenate((boxes, np.full(shape=(boxes.shape[0], 1), fill_value=1., dtype=np.float32)), axis=-1) else: