PASCAL VOC Data Set analysis

Pascal VOC Data Set Analysis Pascal VOC provides a set of standardized and excellent datasets for image recognition and classification, and an image recognition challenge is held every year from 2005 to 2012. This paper mainly analyzes the content of Pascal VOC data set and object recognition in image.
Here we use Pascal VOC2012 as an example. Download Address: Click to open the link. (The system environment in this article is ubuntu14.04) after downloading, you can see the following files in the VOC2012 in the Vocdevkit directory:
Annotations, Imagesets and jpegimages are the most necessary things to know about image object recognition.
The ①jpegimages jpegimages folder contains all the picture information provided by Pascal VOC, including training pictures and test images. These images are named in the "year _ number. jpg" format. The pixel size of the picture is different, but the size of the landscape is around 500*375, the size of the longitudinal figure around 375*500, the basic will not deviate more than 100. (In the following training, the first step is to resize these images to 300*300 or 500*500, all the original images can not be too far from the standard.) These images are the image data used for training and testing validation.
②annotations
The annotations folder contains a label file in XML format, each of which corresponds to a picture in the Jpegimages folder. The exact format of the XML file is as follows: (for 2007_000392.jpg)

<annotation> <folder>VOC2012</folder> <filename>2007_000392.jpg</f                                                           ilename>//File name <source> Image source (unimportant) <database>the VOC2007 database</database> <annotation>pascal voc2007</annotation						
		> <image>flickr</image> </source> <size>//Image size (length width and number of channels) <width>500</width> The corresponding picture is: ③imagesets

Imagesets stores the corresponding image data for each type of challenge. Under Imagesets There are four folders: where action is stored in human actions (such as running, jumping, etc., which is part of the VOC Challenge) layout is stored in the human body parts of the data (human head, hand, Feet and so on, which is also part of the VOC Challenge. The image object recognition data is stored under Main, which is divided into 20 categories. Segmentation is the data that can be used for segmentation.

This is where the main folder is inspected. The main folder contains 20 categories of ***_train.txt, ***_val.txt, and ***_trainval.txt. The contents of these txt are similar to the following: the name of the previous representation image, followed by 1 for the positive sample, 1 for the negative sample. _train is the data that is used for training, and each class has a total of 5,717 train data. The data used in the validation results are stored in the _val, with 5,823 Val data for each class. _trainval the above two together, each class has 11,540. It is necessary to ensure that train and Val do not overlap, that is, training data and validation data can not be duplicated, in the selection of training data should be randomly generated. ④segmentationclass and Segmentationobject

 

These two folders save the object after the segmentation of the image, in the object recognition is not used, here do not do a detailed expansion.

The next thing to look at is how to generate your own training data and test data, as described in the next article.