New Method Of Detecting Objects With High Density And Arbitrary Directionality!

3 main points

✔️ Proposed a highly directional object detection method, DRN
✔️ Proposed data set SKU110K-R for directional and dense object detection
✔️ Improved accuracy in directional and dense object detection

Dynamic Refinement Network for Oriented and Densely Packed Object Detection
written by Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, Changsheng Xu
(Submitted on 20 May 2020 (v1), last revised 10 Jun 2020 (this version, v2))
Comments: Published by CVPR 2020 oral
Subjects: Computer Vision and Pattern Recognition (cs.CV)

Code

Introduction

Object detection has been using Deep learning to produce surprising scores on a number of benchmarks. However, object detection still suffers from one particular problem. That is the challenge of recognizing objects in any direction and detecting them in dense conditions. Inevitably, most detectors were used to train the model with training data, optimize the parameters, and then use a fixed set of parameters. This can lead to a lack of flexibility in specific data during testing.

Modern detection models are based on the RCNN family of frameworks. These methods first generate a large number of horizontal bounding boxes as RoIs and then perform classification and regression on those RoIs. However, for certain objects, the use of such horizontal RoIs can be problematic because of target misalignment. In fact, targets in aerial imagery are usually arbitrarily oriented and dense, which is a problem because of the large number of artifacts that can occur. So another technique is to use anchors to handle targets in arbitrary directions. However, this method is computationally complex because it requires obtaining various anchors such as angle, scale, and aspect ratio. Some researchers have used RoI Trans (which solves the problem of misalignment between the RoI and a directional target by transforming an axially aligned RoI into a rotatable RoI), rotated the RoI itself, and added directionality to the RoI by learning rotationally invariant features. However, there are obvious skewnesses to these. This is because the original purpose of learning is to learn generality from specific data taken from the world of training data. Nevertheless, The relationship between the learned generic model and the addition of only one specific process, target orientation, is distorted. So recently, a method has been proposed that uses dynamic filters and allows for variation for various samples.

The authors improve on this method and propose a Dynamic Refinement Network (DRN) based on CenterNet adding an angle prediction head as a baseline, which is the network of this paper!

Contribution

・We propose a module that adaptively adjusts the receptive field of a neuron based on the shape and orientation of the target. This FSM module is effective in mitigating the dissonance between the receptive field and target.

・We propose two DRHs (for classification and regression tasks, respectively), DRH-C, and DRH-R. These DRHs can be modeled based on the uniqueness and specificity of each sample and refine the predictions for each object.

・A relabeled data set, the SKU110K-R, with accurate annotations of directional bounding boxes, is collected to facilitate the study of directional and dense object detection.

・Consistent accuracy improvements in directionality and dense object detection have been achieved with DOTA, HRSC2016, SKU110K, and SKU110K-R.

Categories related to this article

• Object Detection