[Paper Review] IENet: Interacting Embranchment One Stage Anchor Free Detector for Orientation Aerial Object Detection

Youtian Lin, Pengming Feng+3 more|arXiv (Cornell University)|Dec 2, 2019

Advanced Neural Network ApplicationsComputer Science47 references76 citations

TL;DR

IENet is a one-stage anchor-free detector for orientational aerial object detection that uses a geometric transformation to represent OBBs with HBBs, and an interactive embranchment module with self-attention to fuse features across branches for improved orientation prediction.

ABSTRACT

Object detection in aerial images is a challenging task due to the lack of visible features and variant orientation of objects. Significant progress has been made recently for predicting targets from aerial images with horizontal bounding boxes (HBBs) and oriented bounding boxes (OBBs) using two-stage detectors with region based convolutional neural networks (R-CNN), involving object localization in one stage and object classification in the other. However, the computational complexity in two-stage detectors is often high, especially for orientational object detection, due to anchor matching and using regions of interest (RoI) pooling for feature extraction. In this paper, we propose a one-stage anchor free detector for orientational object detection, namely, an interactive embranchment network (IENet), which is built upon a detector with prediction in per-pixel fashion. First, a novel geometric transformation is employed to better represent the oriented object in angle prediction, then a branch interactive module with a self-attention mechanism is developed to fuse features from classification and box regression branches. Finally, we introduce an enhanced intersection over union (IoU) loss for OBB detection, which is computationally more efficient than regular polygon IoU. Experiments conducted demonstrate the effectiveness and the superiority of our proposed method, as compared with state-of-the-art detectors.

Motivation & Objective

• Motivate efficient orientational object detection in aerial imagery with a one-stage anchor-free approach.
• Develop a representation that converts oriented bounding boxes to surrounding horizontal bounding boxes with orientation parameters.
• Propose an interactive embranchment module with self-attention to fuse classification and regression features for robust orientation regression.

Proposed method

• Adapts FCOS into FCOS-O as a baseline for per-pixel orientation prediction.
• Introduces a geometric transformation that represents an OBB by a surrounding HBB plus orientation parameters [w,h].
• Designs an interactive embranchment (IE) module that fuses features from classification and box regression branches using self-attention.
• Implements an OBB IoU loss derived from an inner box to reduce computation.
• Uses an end-to-end one-stage architecture with an orientational regression branch added to the FCOS-like head.
• Employs a loss combination L = (1/Npos)L_cls + (λ/Npos)L_reg + (ω/Npos)L_ori with specific IoU terms for HBB and OBB.

Experimental results

Research questions

• RQ1 Can an anchor-free one-stage detector achieve competitive performance for oriented aerial object detection?
• RQ2 Does representing OBBs via surrounding HBBs with orientation parameters improve regression stability and accuracy?
• RQ3 Can an IE module with self-attention effectively fuse cross-branch features to improve orientation regression?
• RQ4 Is the proposed OBB IoU loss computationally efficient while maintaining alignment with ground truth OBBs?

Key findings

• IENet outperforms baseline one-stage orientation detectors on public aerial datasets in accuracy and efficiency.
• The geometric transformation enables effective OBB regression by decomposing it into HBB regression and orientation regression.
• The IE module with self-attention improves orientation prediction by leveraging cross-branch feature relationships.
• An inner-box IoU based loss reduces computation compared to full OBB IoU while guiding OBB regression.
• Experiments on DOTA and HRSC2016 show competitive results against state-of-the-art detectors with improved speed and memory efficiency.