luoganttcc的博客

yolov3 中的anchor 的框框是在训练集中聚类所得，在yolov3 中每个格子有9个anchor mask52= [0,1,2] mask26= [3,4,5] mask13= [6,7,8] anchors=[ 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326]yolov3 有9个anchor 点#!/usr/bin/env python3# -*- codin

yolov3 的　loss ，今天终于看完了yolov3-tf2 的源码YOLOv3中Loss部分计算代码 #终点的loss= obj_mask * box_loss_scale*sum((px1-tx1)**2+(px2-tx2)**2) #因为这里有obj_mask的存在，只考虑有目标点的loss xy_loss = obj_mask * box_loss_scale * \ tf.reduce_sum(t

import tensorflow as tftensor = [0, 1, 2, 3] # 1-D examplemask = np.array([True, False, True, False])c=tf.boolean_mask(tensor, mask)print(c)print('------------------------------')tensor = [[1, 2], [3, 4], [5, 6]] # 2-D examplemask = np.array([Tr

最近在看yolov3 的源码，在看yolo_loss的时候遇到了一个卡点，就是将真是标注的box终点坐标转换到anchor点的坐标 true_xy = true_xy * tf.cast(grid_size, tf.float32) - tf.cast(grid, tf.float32) raw_true_xy = y_true[l][..., :2] * grid_shapes[l][:] - gridimport tensorflow as tfimport

#!/usr/bin/env python3# -*- coding: utf-8 -*-"""Created on Thu Jun 17 14:48:19 2021@author: ledi"""from absl import flagsfrom absl.flags import FLAGSimport numpy as npimport tensorflow as tffrom tensorflow.keras import Modelfrom tensorflow.ker

NMS算法的大致过程可以看原文这段话：First, it sorts all detection boxes on the basis of their scores. The detection box M with the maximum score is selected and all other detection boxes with a significant overlap (using a pre-defined threshold) with M are suppressed. T

nms算法就是对同一个anhor 的box 进行基于阈值的去重# python3import numpy as np def py_nms(dets, thresh): """Pure Python NMS baseline.""" #x1、y1、x2、y2、以及score赋值 x1 = dets[:, 0] y1 = dets[:, 1] x2 = dets[:, 2] y2 = dets[:, 3] scores = dets[:, 4]

#!/usr/bin/env python3# -*- coding: utf-8 -*-"""Created on Sun Jun 13 10:24:03 2021@author: ledi"""import pickleimport numpy as npimport tensorflow as tf# data_output = open('output_0.pkl','wb')# pickle.dump(output_0,data_output)# data_output

githubimport tensorflow as tffrom absl.flags import FLAGS@tf.functiondef transform_targets_for_output(y_true, grid_size, anchor_idxs): #这个函数分别对比某一类anchors (一共是三类，每一类对应不同的尺寸的box) #每一类box 对应的尺寸翻倍 # y_true: (N, boxes, (x1, y1, x2, y2, cl

def DarknetConv(x, filters, size, strides=1, batch_norm=True): if strides == 1: padding = 'same' else: x = ZeroPadding2D(((1, 0), (1, 0)))(x) # top left half-padding padding = 'valid' x = Conv2D(filters=filters, kernel_

# import tensorflow as tffrom tensorflow.keras.layers import UpSampling2D,Inputimport numpyfrom tensorflow.keras import Modelx = numpy.array([[1, 2,3], [4, 5,6]])inputs = Input(shape=(2, 3, 1))out =UpSampling2D(size=(4, 4))(inputs)model = Model(in

darknet 是yolov3 里用的模型，如下是是其模型结构#!/usr/bin/env python3# -*- coding: utf-8 -*-"""Created on Mon Jun 7 13:40:34 2021@author: ledi"""import numpy as npimport tensorflow as tffrom tensorflow.keras import Modelfrom tensorflow.keras.layers import (

系统版本：Ubuntu16.04显卡：GTX 1070翻译 + 整理地址: https://pjreddie.com/darknet/install/YOLO: 是实现实时物体检测的系统，Darknet是基于YOLO的框架darknet非常容易安装，它只有２个可选择的依赖：Opencv: 能支持更多格式的图像，并且得到实时的显示GPU: 利用GPU计算，能大大提升YOLO的识别帧率，画面更加流畅安装这两个依赖都必须要先安装基础版yolo—安装基础版yolo—首先将darknet从git

yolo 官网voc数据集下载

git链接python yolo.py

参考链接代码数据集合|python train.py

代码加数据python train.py

１

链接

darknet 链接