Rotating box target detection mmrotate v0.3.1 getting started

1、 Definition of rotating target monitoring

Benefit from the vigorous development of general object detection , At present, most rotating object detection models are based on the classical universal object detection model . With the development of detection task , The horizontal box can no longer meet the needs of researchers in some sub fields . By redefining the object representation , Increase the degree of freedom of regression , Realize rotating rectangle 、 quadrilateral , Even arbitrary shape detection , We call it rotating object detection

2、 What is a rotating box

The most significant difference between rotating object detection and general detection is that the rotating box annotation replaces the horizontal box annotation . They are defined as follows :

Horizontal box: A rectangle with the width along the x-axis and height
along the y-axis. Usually, it can be represented by the coordinates of
2 diagonal vertices (x_i, y_i) (i = 1, 2), or it can be represented by
the coordinates of the center point and the width and height,
(x_center, y_center, width, height).

Rotated box: It is obtained by rotating the horizontal box around the
center point by an angle, and the definition method of its rotated box
is obtained by adding a radian parameter (x_center, y_center, width,
height, theta), where theta = angle * pi / 180. The unit of theta is
rad. When the rotation angle is a multiple of 90°, the rotated box
degenerates into a horizontal box.
The rotated box annotations
exported by the annotation software are usually polygons, which need
to be converted to the rotated box definition method before training

stay MMRotate in , The angle parameter is in radians .

3、Rotation direction（ Direction of rotation ）

Rotate the horizontal box clockwise or counterclockwise around the center point , You can get a rotating box .
Rotate the horizontal box clockwise or counterclockwise around the center point , You can get a rotating box . The rotation direction is closely related to the selection of coordinate system . Right hand coordinate system is used in image space (y, x), among y For the sake of -> Next ,x To the left -> Right . There are two opposite directions of rotation :

3.1 clockwise (CW)

3.2 Anti-clockwise (CCW)

stay MMCV The operators that can set the rotation direction in are :
box_iou_rotated (Defaults to CW)

nms_rotated (Defaults to CW)

RoIAlignRotated (Defaults to CCW)

RiRoIAlignRotated (Defaults to CCW).

MMRotate in , The rotation direction of the rotated box is CW.

3、 Three kinds of shape selection box definitions

because theta The scope of the definition of , In rotating object detection , The following three definitions of rotating boxes have emerged :
For details, please refer to Rotating frame target detection ———— About the definition and solution of the rotation box

1、Doc’ OpenCV Definition

angle∈(0, 90°], theta∈(0, pi / 2], The angle between the width of the rectangle and the positive semi-axis of x is a positive acute angle. This definition comes from the cv2.minAreaRect function in OpenCV, which returns an angle in the range (0, 90°].

2、Dle 135 Long Edge Definition (135°)

: Long Edge Definition (135°),angle∈[-45°, 135°), theta∈[-pi / 4, 3 * pi / 4) and width > height.

3、Dle90

Long Edge Definition (90°),angle∈[-90°, 90°), theta∈[-pi / 2, pi / 2) and width > height

4、MMRotate form

MMRotate consists of 4 main parts, datasets, models, core and apis.

1、 datasets

is for data loading and data augmentation. In this part, we support various datasets for rotated object detection algorithms, useful data augmentation transforms in pipelines for pre-processing image.
For data loading and data enhancement . In this part , We support a variety of datasets for rotating object detection algorithms , This is a useful data enhancement transform in the image preprocessing pipeline .

2、models

contains models and loss functions.
Contains models and loss functions .

3、core

provides evaluation tools for model training and evaluation.
Provide assessment tools for model training and assessment .

4、apis

provides high-level APIs for models training, testing, and inference.
Train for the model 、 Testing and inference provide advanced api.

detecror

neck

rpn head

roi_head

bbox_head

loss

Assigner

MaxConvexIoUAssigner： Assign a corresponding... To each box gt Box or background . Each proposal will be assigned to ’ -1 ', Or a half positive integer to represent the real index of the ground .

-1: Negative samples , Not allocated gt
- Semi positive integer : Positive sample , The distribution of gt The index of ( be based on 0 Start )

Samplers（ Random sampler ）

post_processing（nms）

iou_calculator

Be careful

Due to the different definitions of the rotation box , The following points need to be noted :

Loading annotations

Data augmentation

Assigning samples

Evaluation