1 YoloV3

主要逻辑：1. 特征提取，特征整合（单特征层特征提取和降维）2. 特征融合（大特征层到小特征层）

model.DarkNet

y

* stride_h

WHC transpose to CHW

sigmoid

torch.cat: bs * 3hw * (x+y+w+h+conf+class_conf_list)

view: bs * 3 * h * w * 1-> bs * 3hw * 1

feature_26: torch.Tensor    batch_size(1) * channels(512) * height(26) * width(26)

raw_image: PIL.Image.Imagescale_width: 416scale_height: 416

+ grid_x

* anchor_w

predict_13: torch.Tensor    batch_size(1) * 3(x+y+w+h+conf+class_conf_list) * height(13) * width(13)

w

增加 batch_size 维度： [image]

scaled_image: PIL.Image.Image    width(416) * height(416) * RGB

yolo_utils.letterbox_image

model.YoloNet

predict_52: torch.Tensor    batch_size(1) * 3(x+y+w+h+conf+class_conf_list) * height(52) * width(52)

* anchor_h

model.YoloNet's predict_layer

predict_26: torch.Tensor    batch_size(1) * 3(x+y+w+h+conf+class_conf_list) * height(26) * width(26)

torch.cat

images: torch.Tensor    batch_size(1) * channels(RGB) * height * width

x

主要逻辑：1. 将预测框的 xywh 格式转换为（xmin，ymin，xmax，ymax）的格式2. 将预测结果的（class_conf_list）格式转换为（class_conf，class_label）的格式3. 根据 conf = obj_conf * class_conf 由预设的置信度阈值对预测框进行初步筛选4. 根据 torchvision.ops.nms 由预设的 iou 阈值对预测框进行最终筛选

scaled_image: PIL.Image.Image

as_tensor

predict_layer: torch.Tensor    batch_size(1) * 3(x+y+w+h+conf+class_conf_list) * feature_height * feature_width

tonser 化

predict_output: torch.Tensor    batch_size(1) * anchors_total_num(sum(3hw)) * (x+y+w+h+conf+class_conf_list)

conf

exp

feature_13: torch.Tensor    batch_size(1) * channels(1024) * height(13) * width(13)

* stride_w

h

class_conf_list

yolo_utils.non_max_suppression

yolo_utils.DecodeBox

view: bs * 3 * h * w * 4 -> bs * 3hw * 4

+ grid_y

view: batch_size(1) * 3 * feature_height * feature_width * (x+y+w+h+conf+class_conf_list)

view: bs * 3 * h * w * class_conf_list-> bs * 3hw * class_conf_list

主要逻辑：1. 将输入图片提取五个特征层2. 返回后三个有效的特征层

feature_52: torch.Tensor    batch_size(1) * channels(256) * height(52) * width(52)