全网首发！超全SparseR-CNN实战教程

╰半橙微兮° 2022-08-28 03:53 424阅读 0赞

点击上方“计算机视觉工坊”，选择“星标”

干货第一时间送达

![6c8792345e8dd5464418354793b698d8.png][]

作者‖ 王浩，3D视觉开发者社区签约作者，毕业于北京航空航天大学，人工智能领域优质创作者，CSDN博客认证专家。

编辑‖ 3D视觉开发者社区

# 1、简介 #

目前，目标检测领域中主流的两大类方法。

*  第一大类是 从非Deep时代就被广泛应用的dense detector ，例如DPM，YOLO，RetinaNet，FCOS。在dense detector中，大量的object candidates例如sliding-windows，anchor-boxes， reference-points等被提前预设在图像网格或者特征图网格上，然后直接预测这些candidates到gt的scaling/offest和物体类别。
 *  第二大类是 dense-to-sparse detector ，例如R-CNN家族。这类方法的特点是对一组sparse的candidates预测回归和分类，而这组sparse的candidates来自于dense detector。

但是，dense属性的一些固有局限总让人难以满意，比如：

*  NMS 后处理
 *  many-to-one 正负样本分配
 *  prior candidates的设计

Sparse R-CNN抛弃了anchor boxes或者reference point等dense概念，直接从a sparse set of learnable proposals出发，没有NMS后处理，整个网络异常干净和简洁，可以看做是在dense（单阶段），dense2sparse（二阶段）之外的一个全新的检测范式。

![1f12aad2649db2a09526bbe6883797b3.png][]

# 2、Detectron2介绍 #

**Detectron2** 是Facebook第二代检测工具箱，支持目标检测、实例分割、姿态估计、语义分割和全景分割等任务。

我们使用Detectron2很方便的实现模型的训练、测试以及模型转换。所以现在很多的新模型都是在Detectron2开发。

### 2.1 物体检测的BaseLine（COCO数据集） ###

Faster R-CNN:

<table> 
 <thead> 
 <tr> 
 <td>Name</td> 
 <td>lr sched</td> 
 <td>train time (s/iter)</td> 
 <td>inference time (s/im)</td> 
 <td>train mem (GB)</td> 
 <td>box AP</td> 
 <td>model id</td> 
 <td>download</td> 
 </tr> 
 </thead> 
 <tbody> 
 <tr> 
 <td>R50-C4[1]</td> 
 <td>1x</td> 
 <td>0.551</td> 
 <td>0.110</td> 
 <td>4.8</td> 
 <td>35.7</td> 
 <td>137257644</td> 
 <td>model[2]&nbsp;|&nbsp;metrics[3]</td> 
 </tr> 
 <tr> 
 <td>R50-DC5[4]</td> 
 <td>1x</td> 
 <td>0.380</td> 
 <td>0.068</td> 
 <td>5.0</td> 
 <td>37.3</td> 
 <td>137847829</td> 
 <td>model[5]&nbsp;|&nbsp;metrics[6]</td> 
 </tr> 
 <tr> 
 <td>R50-FPN[7]</td> 
 <td>1x</td> 
 <td>0.210</td> 
 <td>0.055</td> 
 <td>3.0</td> 
 <td>37.9</td> 
 <td>137257794</td> 
 <td>model[8]&nbsp;|&nbsp;metrics[9]</td> 
 </tr> 
 <tr> 
 <td>R50-C4[10]</td> 
 <td>3x</td> 
 <td>0.543</td> 
 <td>0.110</td> 
 <td>4.8</td> 
 <td>38.4</td> 
 <td>137849393</td> 
 <td>model[11]&nbsp;|&nbsp;metrics[12]</td> 
 </tr> 
 <tr> 
 <td>R50-DC5[13]</td> 
 <td>3x</td> 
 <td>0.378</td> 
 <td>0.073</td> 
 <td>5.0</td> 
 <td>39.0</td> 
 <td>137849425</td> 
 <td>model[14]&nbsp;|&nbsp;metrics[15]</td> 
 </tr> 
 <tr> 
 <td>R50-FPN[16]</td> 
 <td>3x</td> 
 <td>0.209</td> 
 <td>0.047</td> 
 <td>3.0</td> 
 <td>40.2</td> 
 <td>137849458</td> 
 <td>model[17]&nbsp;|&nbsp;metrics[18]</td> 
 </tr> 
 <tr> 
 <td>R101-C4[19]</td> 
 <td>3x</td> 
 <td>0.619</td> 
 <td>0.149</td> 
 <td>5.9</td> 
 <td>41.1</td> 
 <td>138204752</td> 
 <td>model[20]&nbsp;|&nbsp;metrics[21]</td> 
 </tr> 
 <tr> 
 <td>R101-DC5[22]</td> 
 <td>3x</td> 
 <td>0.452</td> 
 <td>0.082</td> 
 <td>6.1</td> 
 <td>40.6</td> 
 <td>138204841</td> 
 <td>model[23]&nbsp;|&nbsp;metrics[24]</td> 
 </tr> 
 <tr> 
 <td>R101-FPN[25]</td> 
 <td>3x</td> 
 <td>0.286</td> 
 <td>0.063</td> 
 <td>4.1</td> 
 <td>42.0</td> 
 <td>137851257</td> 
 <td>model[26]&nbsp;|&nbsp;metrics[27]</td> 
 </tr> 
 <tr> 
 <td>X101-FPN[28]</td> 
 <td>3x</td> 
 <td>0.638</td> 
 <td>0.120</td> 
 <td>6.7</td> 
 <td>43.0</td> 
 <td>139173657</td> 
 <td>model[29]&nbsp;|&nbsp;metrics[30]</td> 
 </tr> 
 </tbody> 
</table>

RetinaNet:

RPN & Fast R-CNN:

<table> 
 <thead> 
 <tr> 
 <td>Name</td> 
 <td>lr sched</td> 
 <td>train time (s/iter)</td> 
 <td>inference time (s/im)</td> 
 <td>train mem (GB)</td> 
 <td>box AP</td> 
 <td>prop. AR</td> 
 <td>model id</td> 
 <td>download</td> 
 </tr> 
 </thead> 
 <tbody> 
 <tr> 
 <td>RPN R50-C4[40]</td> 
 <td>1x</td> 
 <td>0.130</td> 
 <td>0.051</td> 
 <td>1.5</td> 
 <td> </td> 
 <td>51.6</td> 
 <td>137258005</td> 
 <td>model[41]&nbsp;|&nbsp;metrics[42]</td> 
 </tr> 
 <tr> 
 <td>RPN R50-FPN[43]</td> 
 <td>1x</td> 
 <td>0.186</td> 
 <td>0.045</td> 
 <td>2.7</td> 
 <td> </td> 
 <td>58.0</td> 
 <td>137258492</td> 
 <td>model[44]&nbsp;|&nbsp;metrics[45]</td> 
 </tr> 
 <tr> 
 <td>Fast R-CNN R50-FPN[46]</td> 
 <td>1x</td> 
 <td>0.140</td> 
 <td>0.035</td> 
 <td>2.6</td> 
 <td>37.8</td> 
 <td> </td> 
 <td>137635226</td> 
 <td>model[47]&nbsp;|&nbsp;metrics[48]</td> 
 </tr> 
 </tbody> 
</table>

### 2.2 使用 Mask R-CNN 的 COCO 实例分割BaseLine ###

<table> 
 <thead> 
 <tr> 
 <td>Name</td> 
 <td>lr sched</td> 
 <td>train time (s/iter)</td> 
 <td>inference time (s/im)</td> 
 <td>train mem (GB)</td> 
 <td>box AP</td> 
 <td>mask AP</td> 
 <td>model id</td> 
 <td>download</td> 
 </tr> 
 </thead> 
 <tbody> 
 <tr> 
 <td>R50-C4[49]</td> 
 <td>1x</td> 
 <td>0.584</td> 
 <td>0.117</td> 
 <td>5.2</td> 
 <td>36.8</td> 
 <td>32.2</td> 
 <td>137259246</td> 
 <td>model[50]&nbsp;|&nbsp;metrics[51]</td> 
 </tr> 
 <tr> 
 <td>R50-DC5[52]</td> 
 <td>1x</td> 
 <td>0.471</td> 
 <td>0.074</td> 
 <td>6.5</td> 
 <td>38.3</td> 
 <td>34.2</td> 
 <td>137260150</td> 
 <td>model[53]&nbsp;|&nbsp;metrics[54]</td> 
 </tr> 
 <tr> 
 <td>R50-FPN[55]</td> 
 <td>1x</td> 
 <td>0.261</td> 
 <td>0.053</td> 
 <td>3.4</td> 
 <td>38.6</td> 
 <td>35.2</td> 
 <td>137260431</td> 
 <td>model[56]&nbsp;|&nbsp;metrics[57]</td> 
 </tr> 
 <tr> 
 <td>R50-C4[58]</td> 
 <td>3x</td> 
 <td>0.575</td> 
 <td>0.118</td> 
 <td>5.2</td> 
 <td>39.8</td> 
 <td>34.4</td> 
 <td>137849525</td> 
 <td>model[59]&nbsp;|&nbsp;metrics[60]</td> 
 </tr> 
 <tr> 
 <td>R50-DC5[61]</td> 
 <td>3x</td> 
 <td>0.470</td> 
 <td>0.075</td> 
 <td>6.5</td> 
 <td>40.0</td> 
 <td>35.9</td> 
 <td>137849551</td> 
 <td>model[62]&nbsp;|&nbsp;metrics[63]</td> 
 </tr> 
 <tr> 
 <td>R50-FPN[64]</td> 
 <td>3x</td> 
 <td>0.261</td> 
 <td>0.055</td> 
 <td>3.4</td> 
 <td>41.0</td> 
 <td>37.2</td> 
 <td>137849600</td> 
 <td>model[65]&nbsp;|&nbsp;metrics[66]</td> 
 </tr> 
 <tr> 
 <td>R101-C4[67]</td> 
 <td>3x</td> 
 <td>0.652</td> 
 <td>0.155</td> 
 <td>6.3</td> 
 <td>42.6</td> 
 <td>36.7</td> 
 <td>138363239</td> 
 <td>model[68]&nbsp;|&nbsp;metrics[69]</td> 
 </tr> 
 <tr> 
 <td>R101-DC5[70]</td> 
 <td>3x</td> 
 <td>0.545</td> 
 <td>0.155</td> 
 <td>7.6</td> 
 <td>41.9</td> 
 <td>37.3</td> 
 <td>138363294</td> 
 <td>model[71]&nbsp;|&nbsp;metrics[72]</td> 
 </tr> 
 <tr> 
 <td>R101-FPN[73]</td> 
 <td>3x</td> 
 <td>0.340</td> 
 <td>0.070</td> 
 <td>4.6</td> 
 <td>42.9</td> 
 <td>38.6</td> 
 <td>138205316</td> 
 <td>model[74]&nbsp;|&nbsp;metrics[75]</td> 
 </tr> 
 <tr> 
 <td>X101-FPN[76]</td> 
 <td>3x</td> 
 <td>0.690</td> 
 <td>0.129</td> 
 <td>7.2</td> 
 <td>44.3</td> 
 <td>39.5</td> 
 <td>139653917</td> 
 <td>model[77]&nbsp;|&nbsp;metrics[78]</td> 
 </tr> 
 </tbody> 
</table>

# 3、搭建Sparse R-CNN测试环境 #

本地环境：操作系统：win10、Cuda11.0。

## 3.1 创建虚拟环境 ##

创建虚拟环境，并激活环境。

conda create --name sparsercnn python=3.7
    activate sparsercnn
    conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0-c pytorch

## 3.2 安装apex库 ##

APEX是英伟达开源的，完美支持PyTorch框架，用于改变数据格式来减小模型显存占用的工具。其中最有价值的是amp（Automatic Mixed Precision），将模型的大部分操作都用Float16数据类型测试，一些特别操作仍然使用Float32。并且用户仅仅通过三行代码即可完美将自己的训练代码迁移到该模型。实验证明，使用Float16作为大部分操作的数据类型，并没有降低参数，在一些实验中，反而由于可以增大Batch size，带来精度上的提升，以及训练速度上的提升。

### 3.2.1 下载apex库 ###

网址 https://github.com/NVIDIA/apex，下载到本地文件夹。解压后进入到apex的目录安装依赖。在执行命令；

cd C:\Users\WH\Downloads\apex-master #进入apex目录
    pip install -r requirements.txt

### 3.2.2 安装apex库 ###

依赖安装完后，打开cmd，cd进入到刚刚下载完的apex-master路径下，运行：

python setup.py install

然后跑了一堆东西，最后是这样的：

![821caacaaadad38322246b2f185d8c9d.png][]

安装完成！

## 3.3 安装fvcore库 ##

fvcore库的简介 fvcore是一个轻量级的核心库，它提供了在各种计算机视觉框架(如Detectron2)中共享的最常见和最基本的功能。这个库基于Python 3.6+和PyTorch。这个库中的所有组件都经过了类型注释、测试和基准测试。Facebook 的人工智能实验室即FAIR的计算机视觉组负责维护这个库。

github地址：https://github.com/facebookresearch/fvcore

执行命令

conda install -c fvcore -c iopath -c conda-forge fvcore

## 3.4 安装Detectron2环境需要其他库 ##

安装pycocotools

pip install pycocotools

安装cv2

pip install opencv-python

安装 antlr4

pip install antlr4-python3-runtime

安装future

pip install future

安装protobuf

pip install protobuf

安装absl

pip install absl-py

安装tensorboard

pip install tensorboard

安装pydot

pip install pydot

安装scipy

pip install scipy -i https://pypi.tuna.tsinghua.edu.cn/simple

## 3.5 编译Sparse R-CNN ##

进入Sparse R-CNN目录，目录根据自己的实际情况更改。

cd D:\SparseR-CNN-main

编译Detectron2和Sparse R-CNN

python setup.py build develop

看到如下信息，则表明安装完成，如果缺少库的情况，则需要安装库，再编译，直到编译成功！

![3f918c30cfe43a1c2e9cb1d5d30eece0.png][]

# 4、测试环境 #

新建input\_img和output\_img文件夹，imgs文件夹存放待测试的图片。

图片如下：

![a9bcd2a4ec4bf3e19c240237d5204890.png][]

下载模型“r50\_100pro\_3x\_model.pth”，将其放到“projects/SparseRCNN”目录下面。

执行命令：

python demo/demo.py --config-file projects/SparseRCNN/configs/sparsercnn.res50.100pro.3x.yaml --input imgs/*.jpg --output imgout --opts MODEL.WEIGHTS projects/SparseRCNN/r50_100pro_3x_model.pth

运行结果：

![c63d24f0cbbc6aeebaeec3a6333d96aa.png][]

能够运行demo说明环境已经没有问题了。

# 5、制作数据集 #

本次采用的数据集是Labelme标注的数据集，地址：链接：https://pan.baidu.com/s/1nxo9-NpNWKK4PwDZqwKxGQ 提取码：kp4e，需要将其转为COCO格式的数据集。转换代码如下：

新建labelme2coco.py

import argparse
    import json
    import matplotlib.pyplot as plt
    import skimage.io as io
    import cv2
    from labelme import utils
    import numpy as np
    import glob
    import PIL.Image
    
    
    REQUIRE_MASK = False
    labels = {'aircraft': 1, 'oiltank': 2}
    
    
    
    
    class labelme2coco(object):
        def __init__(self, labelme_json=[], save_json_path='./new.json'):
            '''
            :param labelme_json: the list of all labelme json file paths
            :param save_json_path: the path to save new json
            '''
            self.labelme_json = labelme_json
            self.save_json_path = save_json_path
            self.images = []
            self.categories = []
            self.annotations = []
            # self.data_coco = {}
            self.label = []
            self.annID = 1
            self.height = 0
            self.width = 0
            self.require_mask = REQUIRE_MASK
            self.save_json()
    
    
        def data_transfer(self):
            for num, json_file in enumerate(self.labelme_json):
                if not json_file == self.save_json_path:
                    with open(json_file, 'r') as fp:
                        data = json.load(fp)
                        self.images.append(self.image(data, num))
                        for shapes in data['shapes']:
                            print("label is ")
                            print(shapes['label'])
                            label = shapes['label']
                            #                    if label[1] not in self.label:
                            if label not in self.label:
                                print("find new category: ")
                                self.categories.append(self.categorie(label))
                                print(self.categories)
                                # self.label.append(label[1])
                                self.label.append(label)
                            points = shapes['points']
                            self.annotations.append(self.annotation(points, label, num))
                            self.annID += 1
    
    
        def image(self, data, num):
            image = {}
            img = utils.img_b64_to_arr(data['imageData'])
            height, width = img.shape[:2]
            img = None
            image['height'] = height
            image['width'] = width
            image['id'] = num + 1
            image['file_name'] = data['imagePath'].split('/')[-1]
            self.height = height
            self.width = width
            return image
    
    
        def categorie(self, label):
            categorie = {}
            categorie['supercategory'] = label
            #        categorie['supercategory'] = label
            categorie['id'] = labels[label]  # 0 默认为背景
            categorie['name'] = label
            return categorie
    
    
        def annotation(self, points, label, num):
            annotation = {}
            print(points)
            x1 = points[0][0]
            y1 = points[0][1]
            x2 = points[1][0]
            y2 = points[1][1]
            contour = np.array([[x1, y1], [x2, y1], [x2, y2], [x1, y2]])  # points = [[x1, y1], [x2, y2]] for rectangle
            contour = contour.astype(int)
            area = cv2.contourArea(contour)
            print("contour is ", contour, " area = ", area)
            annotation['segmentation'] = [list(np.asarray([[x1, y1], [x2, y1], [x2, y2], [x1, y2]]).flatten())]
            # [list(np.asarray(contour).flatten())]
            annotation['iscrowd'] = 0
            annotation['area'] = area
            annotation['image_id'] = num + 1
    
    
            if self.require_mask:
                annotation['bbox'] = list(map(float, self.getbbox(points)))
            else:
                x1 = points[0][0]
                y1 = points[0][1]
                width = points[1][0] - x1
                height = points[1][1] - y1
                annotation['bbox'] = list(np.asarray([x1, y1, width, height]).flatten())
    
    
            annotation['category_id'] = self.getcatid(label)
            annotation['id'] = self.annID
            return annotation
    
    
        def getcatid(self, label):
            for categorie in self.categories:
                #            if label[1]==categorie['name']:
                if label == categorie['name']:
                    return categorie['id']
            return -1
    
    
        def getbbox(self, points):
            polygons = points
            mask = self.polygons_to_mask([self.height, self.width], polygons)
            return self.mask2box(mask)
    
    
        def mask2box(self, mask):
    
    
            # np.where(mask==1)
            index = np.argwhere(mask == 1)
            rows = index[:, 0]
            clos = index[:, 1]
    
    
            left_top_r = np.min(rows)  # y
            left_top_c = np.min(clos)  # x
            right_bottom_r = np.max(rows)
            right_bottom_c = np.max(clos)
            return [left_top_c, left_top_r, right_bottom_c - left_top_c, right_bottom_r - left_top_r]
    
    
        def polygons_to_mask(self, img_shape, polygons):
            mask = np.zeros(img_shape, dtype=np.uint8)
            mask = PIL.Image.fromarray(mask)
            xy = list(map(tuple, polygons))
            PIL.ImageDraw.Draw(mask).polygon(xy=xy, outline=1, fill=1)
            mask = np.array(mask, dtype=bool)
            return mask
    
    
        def data2coco(self):
            data_coco = {}
            data_coco['images'] = self.images
            data_coco['categories'] = self.categories
            data_coco['annotations'] = self.annotations
            return data_coco
    
    
        def save_json(self):
            print("in save_json")
            self.data_transfer()
            self.data_coco = self.data2coco()
    
    
            print(self.save_json_path)
            json.dump(self.data_coco, open(self.save_json_path, 'w'), indent=4)
    
    
    
    
    labelme_json = glob.glob('LabelmeData/*.json')
    from sklearn.model_selection import train_test_split
    
    
    trainval_files, test_files = train_test_split(labelme_json, test_size=0.2, random_state=55)
    import os
    
    
    if not os.path.exists("projects/SparseRCNN/datasets/coco/annotations"):
        os.makedirs("projects/SparseRCNN/datasets/coco/annotations/")
    if not os.path.exists("projects/SparseRCNN/datasets/coco/train2017"):
        os.makedirs("projects/SparseRCNN/datasets/coco/train2017")
    if not os.path.exists("projects/SparseRCNN/datasets/coco/val2017"):
        os.makedirs("projects/SparseRCNN/datasets/coco/val2017")
    labelme2coco(trainval_files, 'projects/SparseRCNN/datasets/coco/annotations/instances_train2017.json')
    labelme2coco(test_files, 'projects/SparseRCNN/datasets/coco/annotations/instances_val2017.json')
    import shutil
    
    
    for file in trainval_files:
        shutil.copy(os.path.splitext(file)[0] + ".jpg", "projects/SparseRCNN/datasets/coco/train2017/")
    for file in test_files:
        shutil.copy(os.path.splitext(file)[0] + ".jpg", "projects/SparseRCNN/datasets/coco/val2017/")

# 6、配置训练环境 #

## 6.1 更改预训练模型的size ##

在projects/SparseRCNN目录，新建change\_model\_size.py文件

import torch
    import numpy as np
    import pickle
    num_class = 2
    pretrained_weights  = torch.load('r50_100pro_3x_model.pth')
    
    
    pretrained_weights["head.head_series.0.class_logits.weight"].resize_(num_class,256)
    pretrained_weights["head.head_series.0.class_logits.bias"].resize_(num_class)
    pretrained_weights["head.head_series.1.class_logits.weight"].resize_(num_class,256)
    pretrained_weights["head.head_series.1.class_logits.bias"].resize_(num_class)
    pretrained_weights["head.head_series.2.class_logits.weight"].resize_(num_class,256)
    pretrained_weights["head.head_series.2.class_logits.bias"].resize_(num_class)
    pretrained_weights["head.head_series.3.class_logits.weight"].resize_(num_class,256)
    pretrained_weights["head.head_series.3.class_logits.bias"].resize_(num_class)
    pretrained_weights["head.head_series.4.class_logits.weight"].resize_(num_class,256)
    pretrained_weights["head.head_series.4.class_logits.bias"].resize_(num_class)
    pretrained_weights["head.head_series.5.class_logits.weight"].resize_(num_class,256)
    pretrained_weights["head.head_series.5.class_logits.bias"].resize_(num_class)
    torch.save(pretrained_weights, "model_%d.pth"%num_class)

这个文件的目的是修改模型输出的size，numclass按照本次打算训练的数据集的类别设置。

## 6.2 修改config参数 ##

路径：“detectron2/engine/defaults.py”

\--config-file：模型的配置文件，SparseRCNN的模型配置文件放在“projects/SparseRCNN/configs”下面。名字和预训练模型对应。

parser.add_argument("--config-file", default="./configs/sparsercnn.res50.100pro.3x.yaml", metavar="FILE", help="path to config file")

resume 是否再次，训练，如果设置为true，则接着上次训练的结果训练。所以第一次训练不用设置。

parser.add_argument(
    "--resume",
        action="store_true",
        help="Whether to attempt to resume from the checkpoint directory. "
    "See documentation of `DefaultTrainer.resume_or_load()` for what it means.",
    )

\--num-gpus，gpu的个数，如果只有一个设置为1，如果有多个，可以自己设置想用的个数。

parser.add_argument("--num-gpus", type=int, default=1, help="number of gpus *per machine*")

opts指的是yaml文件的参数。

上面的参数可以设置，也可以不设置，设置之后可以直接运行不用再考虑设置参数，如果不设置每次训练的时候配置一次参数。

修改类别，文件路径“projects/SparseRCNN/config.py”，

cfg.MODEL.SparseRCNN.NUM_CLASSES = 2

![6385ca47a301c083d278e04bcb7a71c1.png][]

修改yaml文件参数

sparsercnn.res50.100pro.3x.yaml中修改预训练模型的路径。

WEIGHTS: "model_2.pth"

BASE\_LR：设置学习率。

STEPS：设置训练多少步之后调整学习率。

MAX\_ITER：最大迭代次数。

CHECKPOINT\_PERIOD：设置迭代多少次保存一次模型

IMS\_PER\_BATCH：batchsize的大小，根据显存大小设置。

NUM\_CLASSES：数据集中物体类别的种类。

NUM\_PROPOSALS：提议框的个数。

BASE_LR: 0.00025#在Base-SparseRCNN.yaml中
    IMS_PER_BATCH: 2#在Base-SparseRCNN.yaml中
    NUM_CLASSES：2
    STEPS: (21000, 25000)
    MAX_ITER: 54000
    CHECKPOINT_PERIOD: 5000

## 6.3 修改train\_net.py ##

主要修改该setup函数，增加数据集注册。

NUM_CLASSES=2
    def setup(args):
    """
    Create configs and perform basic setups.
    """
        register_coco_instances("train", {}, "datasets/coco/annotations/instances_train2017.json",
    "datasets/coco/train2017")
        register_coco_instances("test", {}, "datasets/coco/annotations/instances_val2017.json",
    "datasets/coco/val2017")
    
    
        cfg = get_cfg()
        add_sparsercnn_config(cfg)
        cfg.merge_from_file(args.config_file)
        cfg.merge_from_list(args.opts)
        cfg.DATASETS.TRAIN = ("train",)
        cfg.DATASETS.TEST = ("test",)
        cfg.MODEL.SparseRCNN.NUM_CLASSES = NUM_CLASSES
        cfg.MODEL.ROI_HEADS.NUM_CLASSES=NUM_CLASSES
        cfg.freeze()
        default_setup(cfg, args)
    return cfg

还要修改detectron2/engine/launch.py，在launch函数下面增加一句

dist.init_process_group('gloo', init_method='file://tmp/somefile', rank=0, world_size=1)

如下图：

![e3dfad79104a3ea0928fb3bc25df5b9f.png][]

这句话的作用是初始化分布式训练，因为我们没有使用分布式，所以没有初始化，但是不初始化就会报错，所以加上这句。

# 7、训练 #

两种启动方式：

第一种，命令行：进入“projects/SparseRCNN/”目录下,执行：

python train_net.py

第二种，直接在pycharm 直接运行train\_net.py.

训练结果：

![b7fba4ab9416aa38dfa0409d2f2cf474.png][]

从训练结果上看，效果确实不错，和CenterNet2的结果相差不大，不过模型很大，大约有1.2G，比CenterNet2的模型大了一倍多。

![7d80b37efebc4b5f53c759dd786ffe69.png][]

# 8、测试 #

修改demo/demo.py

## 8.1 修改setup\_cfg函数 ##

![21134e12bf48143d6c950cb44974d286.png][]

在红框的位置增加代码，详细如下面的代码。

NUM_CLASSES=2
    def setup_cfg(args):
    # load config from file and command-line arguments
        cfg = get_cfg()
    from projects.SparseRCNN.sparsercnn import add_sparsercnn_config
        add_sparsercnn_config(cfg)
        cfg.MODEL.SparseRCNN.NUM_CLASSES = NUM_CLASSES
        cfg.MODEL.ROI_HEADS.NUM_CLASSES = NUM_CLASSES
        cfg.merge_from_file(args.config_file)
        cfg.merge_from_list(args.opts)
    # Set score_threshold for builtin models
        cfg.MODEL.RETINANET.SCORE_THRESH_TEST = args.confidence_threshold
        cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = args.confidence_threshold
        cfg.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = args.confidence_threshold
        cfg.freeze()
    return cfg

## 8.2 修改显示类别 ##

在demo/predictor.py

![72f6453bbce47bd4e2626d69e0f5deff.png][]

代码：

visualizer.metadata.thing_classes[:10] = ["aircraft", "oiltank"]

然后进入SparseR-CNN-main目录，执行如下命令：

python demo/demo.py --config-file projects/SparseRCNN/configs/sparsercnn.res50.100pro.3x.yaml --input img/*jpg --output out --opts MODEL.WEIGHTS projects/Spa
    rseRCNN/output/model_final.pth

运行结果：

![f044510192d9ae02d148f8f43052c5ed.png][]

### References ###

`[1]` R50-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_50\_C4\_1x.yaml*  
`[2]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_C4\_1x/137257644/model\_final\_721ade.pkl*  
`[3]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_C4\_1x/137257644/metrics.json*  
`[4]` R50-DC5: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_50\_DC5\_1x.yaml*  
`[5]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_DC5\_1x/137847829/model\_final\_51d356.pkl*  
`[6]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_DC5\_1x/137847829/metrics.json*  
`[7]` R50-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_50\_FPN\_1x.yaml*  
`[8]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_FPN\_1x/137257794/model\_final\_b275ba.pkl*  
`[9]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_FPN\_1x/137257794/metrics.json*  
`[10]` R50-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_50\_C4\_3x.yaml*  
`[11]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_C4\_3x/137849393/model\_final\_f97cb7.pkl*  
`[12]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_C4\_3x/137849393/metrics.json*  
`[13]` R50-DC5: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_50\_DC5\_3x.yaml*  
`[14]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_DC5\_3x/137849425/model\_final\_68d202.pkl*  
`[15]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_DC5\_3x/137849425/metrics.json*  
`[16]` R50-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_50\_FPN\_3x.yaml*  
`[17]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_FPN\_3x/137849458/model\_final\_280758.pkl*  
`[18]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_50\_FPN\_3x/137849458/metrics.json*  
`[19]` R101-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_101\_C4\_3x.yaml*  
`[20]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_101\_C4\_3x/138204752/model\_final\_298dad.pkl*  
`[21]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_101\_C4\_3x/138204752/metrics.json*  
`[22]` R101-DC5: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_101\_DC5\_3x.yaml*  
`[23]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_101\_DC5\_3x/138204841/model\_final\_3e0943.pkl*  
`[24]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_101\_DC5\_3x/138204841/metrics.json*  
`[25]` R101-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_R\_101\_FPN\_3x.yaml*  
`[26]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_101\_FPN\_3x/137851257/model\_final\_f6e8b1.pkl*  
`[27]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_R\_101\_FPN\_3x/137851257/metrics.json*  
`[28]` X101-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/faster\_rcnn\_X\_101\_32x8d\_FPN\_3x.yaml*  
`[29]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_X\_101\_32x8d\_FPN\_3x/139173657/model\_final\_68b088.pkl*  
`[30]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster\_rcnn\_X\_101\_32x8d\_FPN\_3x/139173657/metrics.json*  
`[31]` R50: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/retinanet\_R\_50\_FPN\_1x.yaml*  
`[32]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet\_R\_50\_FPN\_1x/137593951/model\_final\_b796dc.pkl*  
`[33]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet\_R\_50\_FPN\_1x/137593951/metrics.json*  
`[34]` R50: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/retinanet\_R\_50\_FPN\_3x.yaml*  
`[35]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet\_R\_50\_FPN\_3x/137849486/model\_final\_4cafe0.pkl*  
`[36]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet\_R\_50\_FPN\_3x/137849486/metrics.json*  
`[37]` R101: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/retinanet\_R\_101\_FPN\_3x.yaml*  
`[38]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet\_R\_101\_FPN\_3x/138363263/model\_final\_59f53c.pkl*  
`[39]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/retinanet\_R\_101\_FPN\_3x/138363263/metrics.json*  
`[40]` RPN R50-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/rpn\_R\_50\_C4\_1x.yaml*  
`[41]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn\_R\_50\_C4\_1x/137258005/model\_final\_450694.pkl*  
`[42]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn\_R\_50\_C4\_1x/137258005/metrics.json*  
`[43]` RPN R50-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/rpn\_R\_50\_FPN\_1x.yaml*  
`[44]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn\_R\_50\_FPN\_1x/137258492/model\_final\_02ce48.pkl*  
`[45]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/rpn\_R\_50\_FPN\_1x/137258492/metrics.json*  
`[46]` Fast R-CNN R50-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-Detection/fast\_rcnn\_R\_50\_FPN\_1x.yaml*  
`[47]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/fast\_rcnn\_R\_50\_FPN\_1x/137635226/model\_final\_e5f7ce.pkl*  
`[48]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/fast\_rcnn\_R\_50\_FPN\_1x/137635226/metrics.json*  
`[49]` R50-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_C4\_1x.yaml*  
`[50]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_C4\_1x/137259246/model\_final\_9243eb.pkl*  
`[51]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_C4\_1x/137259246/metrics.json*  
`[52]` R50-DC5: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_DC5\_1x.yaml*  
`[53]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_DC5\_1x/137260150/model\_final\_4f86c3.pkl*  
`[54]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_DC5\_1x/137260150/metrics.json*  
`[55]` R50-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_FPN\_1x.yaml*  
`[56]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_FPN\_1x/137260431/model\_final\_a54504.pkl*  
`[57]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_FPN\_1x/137260431/metrics.json*  
`[58]` R50-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_C4\_3x.yaml*  
`[59]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_C4\_3x/137849525/model\_final\_4ce675.pkl*  
`[60]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_C4\_3x/137849525/metrics.json*  
`[61]` R50-DC5: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_DC5\_3x.yaml*  
`[62]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_DC5\_3x/137849551/model\_final\_84107b.pkl*  
`[63]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_DC5\_3x/137849551/metrics.json*  
`[64]` R50-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_FPN\_3x.yaml*  
`[65]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_FPN\_3x/137849600/model\_final\_f10217.pkl*  
`[66]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_50\_FPN\_3x/137849600/metrics.json*  
`[67]` R101-C4: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_C4\_3x.yaml*  
`[68]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_C4\_3x/138363239/model\_final\_a2914c.pkl*  
`[69]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_C4\_3x/138363239/metrics.json*  
`[70]` R101-DC5: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_DC5\_3x.yaml*  
`[71]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_DC5\_3x/138363294/model\_final\_0464b7.pkl*  
`[72]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_DC5\_3x/138363294/metrics.json*  
`[73]` R101-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_FPN\_3x.yaml*  
`[74]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_FPN\_3x/138205316/model\_final\_a3ec72.pkl*  
`[75]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_R\_101\_FPN\_3x/138205316/metrics.json*  
`[76]` X101-FPN: *https://github.com/facebookresearch/detectron2/blob/master/configs/COCO-InstanceSegmentation/mask\_rcnn\_X\_101\_32x8d\_FPN\_3x.yaml*  
`[77]` model: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_X\_101\_32x8d\_FPN\_3x/139653917/model\_final\_2d9806.pkl*  
`[78]` metrics: *https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask\_rcnn\_X\_101\_32x8d\_FPN\_3x/139653917/metrics.json*

本文仅做学术分享，如有侵权，请联系删文。

##  ##

**重磅！计算机视觉工坊****-学习****交流群****已成立**

扫码添加小助手微信，可申请加入3D视觉工坊-学术论文写作与投稿 微信交流群，旨在**交流顶会、顶刊、SCI、EI等写作与投稿事宜。**

**同时**也可申请加入我们的细分方向交流群，目前主要有**ORB-SLAM系列源码学习、****3D视觉**、**CV&深度学习**、**SLAM**、**三维重建**、**点云后处理**、**自动驾驶、CV入门、三维测量、VR/AR、3D人脸识别、医疗影像、缺陷检测、行人重识别、目标跟踪、视觉产品落地、视觉竞赛、车牌识别、硬件选型、深度估计、学术交流、求职交流**等微信群，请扫描下面微信号加群，备注：”研究方向+学校/公司+昵称“，例如：”3D视觉 + 上海交大 + 静静“。请按照格式备注，否则不予通过。添加成功后会根据研究方向邀请进去相关微信群。**原创投稿**也请联系。

![d9875b77834300e4857df778131e9fd8.png][]

▲长按加微信群或投稿

![0249870233606335392b8986741da36b.png][]

▲长按关注公众号

##  ##

**3D视觉从入门到精通知识星球**：针对3D视觉领域的**视频课****程（**[**三维重建系列**][Link 1]**、**[**三维点云系列**][Link 2]**、**[**结构光系列**][Link 3]**、**[**手眼标定**][Link 4]**、**[**相机标定**][Link 5]**、激光/视觉SLAM、自动驾驶等****）、****知识点汇总、入门进阶学习路线、最新paper分享、疑问解答**五个方面进行深耕，更有各类大厂的算法工程人员进行技术指导。与此同时，星球将联合知名企业发布3D视觉相关算法开发岗位以及项目对接信息，打造成集技术与就业为一体的铁杆粉丝聚集区，近4000星球成员为创造更好的AI世界共同进步，知识星球入口：

学习3D视觉核心技术，扫描查看介绍，3天内无条件退款

##  ##

![98cb7a872a12f078a55043f04b665efe.png][]

圈里有高质量教程资料、可答疑解惑、助你高效解决问题

**觉得有用，麻烦给个赞和在看~**

[6c8792345e8dd5464418354793b698d8.png]: /images/20220828/1a876e97b8bf4cb8aea6f4ef270f9d03.png
[1f12aad2649db2a09526bbe6883797b3.png]: /images/20220828/6247b84652f64977afa88afd952454a5.png
[821caacaaadad38322246b2f185d8c9d.png]: /images/20220828/ca2ce5515cf14851a0f7e4646bd24146.png
[3f918c30cfe43a1c2e9cb1d5d30eece0.png]: /images/20220828/3b1d006aeb334cb49129466a9148bacb.png
[a9bcd2a4ec4bf3e19c240237d5204890.png]: /images/20220828/f12e12253eb747e7833ecc68c541a0ed.png
[c63d24f0cbbc6aeebaeec3a6333d96aa.png]: /images/20220828/9e41935829b6479cba2016214ad1335a.png
[6385ca47a301c083d278e04bcb7a71c1.png]: /images/20220828/0b9359b6647743faa8ccdb07d0eaae7f.png
[e3dfad79104a3ea0928fb3bc25df5b9f.png]: /images/20220828/a0d4a747d565410ba6bfc16850c90fc1.png
[b7fba4ab9416aa38dfa0409d2f2cf474.png]: /images/20220828/a986047fa8a34178afe322590570911c.png
[7d80b37efebc4b5f53c759dd786ffe69.png]: /images/20220828/2a8d34a2a79841ffb6417facf2919606.png
[21134e12bf48143d6c950cb44974d286.png]: /images/20220828/29e34c745bce4a80b69f925f0c9f5ed7.png
[72f6453bbce47bd4e2626d69e0f5deff.png]: /images/20220828/b0a8c40fc299475a9a85f144f3dfbf99.png
[f044510192d9ae02d148f8f43052c5ed.png]: /images/20220828/93d1a933f814496f9b727333df7b9116.png
[d9875b77834300e4857df778131e9fd8.png]: /images/20220828/387e5c70dd4943dba57001af988aff2f.png
[0249870233606335392b8986741da36b.png]: /images/20220828/cc980afd530a44a8a3a0ca98cb4fb691.png
[Link 1]: http://mp.weixin.qq.com/s?__biz=MzU1MjY4MTA1MQ==&mid=2247549738&idx=1&sn=88b4aa9e091077f107407bddc8da7acd&chksm=fbfc281ecc8ba10823e11ac0044b5a94920806c83d30d8cb90630a41ecb23ad458106b9b1f58&scene=21#wechat_redirect
[Link 2]: http://mp.weixin.qq.com/s?__biz=MzU1MjY4MTA1MQ==&mid=2247551625&idx=1&sn=4a914cfeb2d251ab25421e1e5de930c0&chksm=fbfc21bdcc8ba8abac371478a6b2b9d143a1b31fa799201c6682d9d0b6a4ed4a7da146b3c536&scene=21#wechat_redirect
[Link 3]: http://mp.weixin.qq.com/s?__biz=MzU1MjY4MTA1MQ==&mid=2247546284&idx=2&sn=d067a98047192a04f21489aca338a5d6&chksm=fbfc1e98cc8b978e5fb6b2afd1a473c89e2861374542421b9972f87a6fdfcd47eb50c443042e&scene=21#wechat_redirect
[Link 4]: http://mp.weixin.qq.com/s?__biz=MzU1MjY4MTA1MQ==&mid=2247554670&idx=2&sn=cf3323178d5d1960476208e6303a8479&chksm=fbfc3d5acc8bb44c184084dcf447a5981d297815d62bef699129338d99a0472c622f96ee71d8&scene=21#wechat_redirect
[Link 5]: http://mp.weixin.qq.com/s?__biz=MzU1MjY4MTA1MQ==&mid=2247486422&idx=1&sn=381ab793959aed0eb6f01019dea522d4&chksm=fbff20e2cc88a9f456b80e23d2482fb319b08e89da071b39597bc51942a6aa7b756a6749856c&scene=21#wechat_redirect
[98cb7a872a12f078a55043f04b665efe.png]: /images/20220828/c83d2c91ff144ca2abc7a1ac9170f188.png