Paper address: https://arxiv.org/abs/2105.11111
Translation: https://blog.csdn.net/songyuc/article/details/128227048
I. Overview
The characteristics of aviation targets:
(1) Non-positive axis alignment and arbitrary orientation
(2) Background noise clutter
The core points proposed in the paper:
(1) An efficient adaptive point learning method is proposed to better capture the geometric information of arbitrary orientation instances.
(2) Efficient quality assessment and sample classification methods are used for adaptive point learning , enabling the capture of non-axis-aligned features from nearby objects or background noise. (3) Three orientation transformation functions
are proposedto enhance orientation-accurate classification and localization. (4) A spatial constraint is introducedto penalize outliers for robust adaptive learning. To sum up, it is to obtain feature information, enhance the accuracy of information, and have appropriate robustness.
Datasets:
DOTA, HRSC, UCAS-AOD and DIOR-R
2. Related work
Rotated object detection:
traditional : Predict the rotation angle of the bounding box (1) convert the angle regression to an angle classification task (2) convert the parameterization of the rotated bounding box into a 2D Gaussian distribution The problem to be solved: the direction estimation of the
angle boundary Discontinuity
work content : From another perspective, a more efficient adaptive point-based representation method is introduced.
Axis-Independent Feature Learning:
Traditional : Oriented upright or axially aligned.
(1) Spatially transform the axis-aligned RoIs and learn an axis-independent representation under the supervision of rotated bounding boxes.
(2) A feature alignment module is designed to alleviate the misalignment between axially aligned convolutional features and arbitrary orientation targets.
(3) A feature selection module is proposed to aggregate axial alignment information obtained from convolution kernels of different sizes, shapes, and orientations, and use a dynamic filter generator for further regression.
(4) Use convex hull representation to learn irregular shapes and arrangements, and avoid feature aliasing through learnable feature adaptation.
Work content : The point set-based method proposed in this paper is to extract key features for non-axially aligned aerial objects.
Sample allocation strategy for target detection:
Traditional : Select the IoU threshold of positive samples, introduce a matching measure to evaluate the degree of spatial alignment based on angle anchor boxes, and use alignment-sensitive loss to strengthen the correlation of classification and orientation.
Open problem : Potentially noisy and difficult examples exist.
Work content : This paper proposes an effective quality assessment and sample assignment method to select positive sample points.
三、Oriented RepPoints
Oriented RepPoints is the model proposed in this paper. It is based on DCN (deformable convolutional network) and RepPoints (point set)
RepPointsde's paper address: https://arxiv.org/abs/1904.11490
(1) introduces differentiable Transformation functions to efficiently learn high-quality adaptive points.
(2) A quality metric is proposed to select high-quality oriented reppoints during training.
(3) Employ spatial constraints to facilitate robust adaptive representation point learning.
(1) Adaptive representation point learning with angle
G is represented as an angle conversion function, R represents the learning representation point set, and OP represents the angle box to.
The purpose of MinAeraRect is to find the rotated rectangle with the smallest area from the set of learned representation points for a rotated object.
NearestGTCorner uses ground truth annotations. For each corner point, a nearest neighbor point is selected from the learned representation point set as the predicted corner point, and the selected corner point is used to construct a quadrilateral as the rotation bounding box.
ConvexHull. Based on the Jarvis March algorithm [7_CFA, 11_JarvisMarchAlgorithm], a rotated instance polygon can be defined by the convex hull of a set of points, which is also used in many edge-based methods.
In this paper, MinAeraRect is used in post-processing to obtain standard rotation rectangle predictions, and the other two functions are used to optimize the learning of adaptive representation points during training. This is because the former is not differentiable.
The framework proposed in this paper consists of two stages. In the initial stage, an adaptive representation point set is generated by extracting target center points (feature map bins). The refinement stage further obtains finer adjustments by minimizing the loss function. The formula is clearly expressed in the original text, so it will not be shown here.
(2) Adaptive point evaluation and allocation
For densely distributed and arbitrarily oriented objects in aerial images, learning high-quality points is important to adaptively capture geometric features due to the lack of direct supervision. To this end, this paper proposes an evaluation and assignment method to measure the quality of learned points, which can assign representative samples of adaptive points as positive samples during the training phase. That is to say, this paper proposes a new model to evaluate high-quality points.
4. Related papers
RepPoints: Point Set Representation for Object Detection: https://arxiv.org/abs/1904.11490
Beyond Bounding-Box: Convex-hull Feature Adaptation for Oriented and Densely Packed Object Detection：https://link.csdn.net/?target=https%3A%2F%2Fopenaccess.thecvf.com%2Fcontent%2FCVPR2021%2Fpapers%2FGuo_Beyond_Bounding-Box_Convex-Hull_Feature_Adaptation_for_Oriented_and_Densely_Packed_CVPR_2021_paper.pdf
On the identification of the convex hull of a finite set of points in the plane：https://link.csdn.net/?target=https%3A%2F%2Fwww.sciencedirect.com%2Fscience%2Farticle%2Fabs%2Fpii%2F0020019073900203
Focal Loss for Dense Object Detection：Focal Loss for Dense Object Detection
A Point Set Generation Network for 3D Object Reconstruction from a Single Image：https://arxiv.org/abs/1612.00603