Target detection notes - overview and common data sets

Overview and common data sets

0. summary

• Want to have Classification of network Related basic knowledge
  
• Target detection is divided into two categories ：One-Stage,Two-Stage

1.Two-Stage: Faster R-CNN

​ 1） Go through special modules Generate candidate boxes （RPN）, Looking for the future as well as Adjust the bounding box （ be based on anchors）

​ 2） Based on the previously generated candidate box Further classify and adjust the bounding box （ be based on proposals）

2.One-Stage: SSD,YOLO

​ be based on anchors Directly classify and adjust the bounding box

One-Stage,Two-Stage contrast ：One-Stage testing Faster ,Two-Stage testing More accurate

1.TF Official API

up See the video for details https://www.bilibili.com/video/BV1KE411E7Ch?spm_id_from=333.999.0.0

2.PASCAL VOC2012 Data sets

• Click on **Download the training/validation data (2GB tar file)** download
• There are mainly 20 Categories , It is divided into 4 Categories:

• Dataset catalog

• Usage method ：

1. Read train.txt Each line of information in the file （ Image name ）
2. stay Annotations Find the corresponding .xml file
3. adopt .xml file You can get picture information （ Width , Height , Target location, etc ）
4. stay JPEGImages Find the corresponding picture under the folder , Load memory .

3. MS COCO Data set and usage

There are Object80 class and stuff91 class ,Object80 Class is stuff91 Class A subset of , Generally, target detection only needs Object80 class , Let's see the usage scenarios .

be relative to PASCAL VOC The effect of pre training is better , But it takes more time .

Coco Data sets Official website ：http://cocodataset.org/

SkyDrive address ： link ：https://pan.baidu.com/s/1_kQBJxB019cXpzfqNz7ANA Extraction code ：m6s5

bilibili up Explain the video ：https://www.bilibili.com/video/BV1TK4y1o78H?spm_id_from=333.999.0.0

Recommended blog links ：https://blog.csdn.net/qq_37541097/article/details/112248194

github link ：https://github.com/WZMIAOMIAO/deep-learning-for-image-processing

For more details, please see the official website document ：http://cocodataset.org/#format-data

Ps: It is meaningless to divide the test set with your own data , Because the data distribution is basically the same

import json
#  see json The data structure of the file 
json_path = "G:/coco2017/annotations/instances_val2017.json"
json_labels = json.load(open(json_path,"r"))
print(json_labels)
# >>>json_label
# |-info: Describe file information 
# |-licenses: No impact information 
# |-image:5000 Two elements correspond to 5000 A picture , Contains picture information 
# |-annotations:3w+ Elements correspond to 3w+ The goal is 
# | |-segmentation: Split information 
# | |-area: area 
# | |-iscrowed: Is it overlapping , Generally, only 0 Training for 
# | |-image_id: picture id
# | |-bbox: Mark box information ( Left upper coordinate , Length and width )
# | |_category_id: Category name 
# |_categories:80 Two elements correspond to 80 Classes 
# PS:categories in id Discontinuous , Pay attention during training .

#  About coco Official tools pycocotools Installation 
# liunx:pip install pycocotools
# windows:pip install pycocotools-windows
import os
from pycocotools.coco import COCO
from PIL import Image,ImageDraw
import matplotlib.pyplot as plt

json_path = "G:/coco2017/annotations/instances_val2017.json"
img_path = "G:/coco2017/val2017"
#  Import coco data 
coco = COCO(annotation_file=json_path)
#  Get the index of all pictures 
ids = list(sorted(coco.imgs.keys()))
print("number of image:%d"%len(ids))
#  Get everything coco Category labels 
coco_classes = dict([(v["id"], v["name"]) for k,v in coco.cats.items()])
#  Traverse the first three pictures 
for img_id in ids[:3]:
    #  Obtain the corresponding image id All of the annotations ids Information 
    ann_ids = coco.getAnnIds(img_id)
    #  according to annotations idx Information get all annotation information 
    targets = coco.loadAnns(ann_ids)
    #  Get the name of the image file 
    path = coco.loadImgs(img_id)[0]["file_name"]
    #  Read the image according to the path , Convert to RGB Format 
    img = Image.open(os.path.join(img_path,path)).convert('RGB')
    #  Drawing pictures 
    draw = ImageDraw.Draw(img)
    #  Draw the target box 
    for target in targets:
        x,y,w,h = target["bbox"]  #  Extract the target box information 
        x1,y1,x2,y2 = x,y,int(x+w),int(y+h)  #  Calculate the coordinate 
        draw.rectangle((x1,y1,x2,y2))  #  Draw the target box 
        draw.text((x1,y1),coco_classes[target["category_id"]])  #  Drawing class name 
    #  display picture 
    plt.imshow(img)
    plt.show()

# coco File form of target detection results in 
''' [{ "image_id" :int, #  Record the image to which the target belongs id "category_id" :int, #  Record the category index that predicts the target  "bbox" :[x,y,w,h] #  Record the bounding box information that predicts the target  "score" :float #  Record the probability of predicting the goal  }] '''
#  The output of training results is saved as json form , Compare with the results on the validation set 
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
#  Load the dimension file of the validation set 
coco_ture = COCO(annotation_file="./val_2017.json")
#  Load network at coco2017 Verify the prediction results on the set 
coco_pre = coco_ture.loadRes('./predict_results.json')
#  Output the effect of prediction , indicators mAP, See section 5 Section 
coco_evaluator = COCOeval(cocoGt=coco_ture, cocoDt=coco_pre, iouType="bbox")
coco_evaluator.evaluate()
coco_evaluator.accumulate()
coco_evaluator.summarize()

4. Create your own data set LabelImg&Labelme

LabelImg Generated are xml File with the VOC2012 Agreement

Labelme It's not just the sample that can mark the target detection , Other sample annotations such as semantic segmentation can also be carried out .

• Download and install ：pip install labelImg
• Input labelImg Open software

5. object detection coco Common indicators

TP(True Positive):IOU>0.5 Number of detection frames ( same Ground Truth Calculate only once ).

FP(Flase Positive):IOU<=0.5 The detection frame of ( Or the same... Is detected Ground Truth Redundant detection box ) The number of .

FN(False Negative): Not detected Ground Truth The number of .

Precision( Precision rate ):TP/(TP+FP) Of all the goals predicted by the model , Predict the right proportion .

Recall( Recall rate ):TP/(TP+FN) Among all real targets , The model predicts the correct target proportion .

AP(Average Precision): P-R(Precision-Recall) Area under curve .

mAP(mean Average Precision): That is, each category AP Average value