非结构化环境中的实例级物体检测Instance object detection in unstructured environment-仪器光电学院测控系B704实验室

2、室内场景生成图像训练数据库BigBIRD-GI Dataset

BGID (BigBIRD Generational Indoor Dataset)

当前用于实例级物体检测的常用开源标准训练数据库为厨房场景下开源标准数据集GMU(George Mason University Kitchen Dataset, 简称GMU)。该数据库采用了实例物体开源数据集BigBIRD(Berkeley Instance Recognition Dataset, 简称BigBIRD)中的11个实例物体进行厨房场景下各个实例物体分布情况的图像拍摄，共计6439张图像，并提供了厨房场景下各个实例物体位置框的标注文件。

At present, the common open source standard training database for instance object detection is the George Mason University Kitchen Dataset (GMU). This database uses 11 instance objects that belong to the Berkeley Instance Recognition Dataset (BigBIRD), and then take pictures of these instance objects in nine kinds of kitchen scene, totaling 6439 images with the annotated location bbox files.

厨房场景下开源标准数据集GMU的网址链接如下：

https://cs.gmu.edu/~robot/gmu-kitchens.html

实例物体开源数据集BigBIRD的网址链接如下：

http://rll.berkeley.edu/bigbird/

The open source dataset GMU has the following links:

Https://cs.gmu.edu/~robot/gmu-kitchens.html

The open source standard dataset BigBIRD has the following links:

http://rll.berkeley.edu/bigbird/

由于数据集BigBIRD为每个实例物体仅提供了600张不同视角图像，为进一步丰富实例物体的视角信息与添加多种数据增强变换的多样性，并解决直接在复杂空间场景下进行实例物体多视角图像的采集与标注工作的耗时耗力问题，本实验室采用生成式模型搭建室内场景生成图像训练数据集BigBIRD-GI，所有数据如附件中压缩文件，数据库中的样本图像如图2所示。

Because BigBIRD provides only 600 different viewpoints images for each instance object, in order to further enrich the viewpoint information as well as add the variety of data augmentation transformations, and solve the time-consuming problem of collecting and labeling instance object directly in complex space scenes, we use generative model to build BigBIRD Generational Indoor Dataset (BigBIRD-GI). All the data in BigBIRD-GI are compressed files in the attachment. The sample image in the database is shown in Figure 2.

图2 室内场景生成图像训练数据集BigBIRD-GI部分图像示意图

Fig.2 Some examples in BigBIRD-GI

根据对实例物体在三维空间中的多视角姿态与数据增强变换（位移、尺寸、旋转、拉伸、亮度、饱和度）图像的不同扩增丰富程度，室内场景生成图像训练数据集BigBIRD-GI共包含4种不同的数量等级，其细节构成如下所示：

According to the different extensions level of multi-viewpoints in three-dimensional space images and data augmentation transformations (shift, size, rotation, stretching, brightness, saturation) images of instance objects, the BigBIRD-GI contains four different levels of magnitude. The details of the BigBIRD-GI are as follows:

BigBIRD-GI1-4：使用生成式反卷积网络GDDNE分别为数据集BigBIRD中的每个实例物体自动插值生成8000、6000、22800、88800张不同视角图像，四种不同情况的视角分布分别为：平面内0-360度之间每间隔0.4度、0.6度、0.3度、0.15度获取一张关键帧图像共计900个、600个、1200个、2400个不同旋转角度，以及摄像机在空间内0-90度之间每间隔10度、10度、5度、2.5度获取一张关键帧图像共计10个、10个、19个、37个不同俯角角度。同时在BigBIRD-GI1-3与BigBIRD-GI1-4的训练集中使用生成式反卷积网络GDDNE分别为每个实例物体自动生成6种数据增强变换图像，以及1-2种数据增强叠加变换的21种变换图像，其中每种数据增强图像生成1000张，共计6000张与21000张生成图像两种情况。之后按照2:1:1的复杂背景图像比例，在每张图像上分别放置3个、5个、7个实例物体不同视角图像，同时对有遮挡与无遮挡的样本数量进行等比例设置，并加入500张真实样本数据，共计4500张、10500张、20500张、40500张训练样本数据。

BigBIRD-GI1-4: Using the generative deconvolutional network GDDNE to generate 8000、6000、22800 and 88800 different viewpoint images for each instance object in BigBIRD Dataset automatically. The angle distributions of such four different situations are as follows: 1) A total of 900, 600, 1200, 2400 different rotation angles between 0-360 degrees in the plane at 0.4 degree, 0.6 degree, 0.3 degree, 0.15 degree interval. 2)A total of 10, 10, 19, 37 different depression angles are set for the camera in space, each at intervals of 10, 10, 5, 2.5 degrees between 0-90 degrees in space. At the same time, the GDDNE is used to automatically generate six kinds of data augmentation transformations images for each instance object in BIGBIRD-GI-3, in another case, 1-2 kinds of superposition transformations are carried out in 6 kinds of transformations, totaling 21 kinds of transformationsimages in BIGBIRD-GI-4. Each kind of data augmentation images are 1000, totally 6000 and 21000 generated images. Then, according to the ratio of 2:1:1 complex background image, three, five and seven different instance objects are placed on each image. At the same time, the number of occluded and unshielded samples is set in equal proportion. Finally, 400 real sample data are added, totaling 1400, 2400, 4400 and 8400 training sample data.