Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Anchor-free Oriented Proposal Generator for Object Detection (2110.01931v2)

Published 5 Oct 2021 in cs.CV

Abstract: Oriented object detection is a practical and challenging task in remote sensing image interpretation. Nowadays, oriented detectors mostly use horizontal boxes as intermedium to derive oriented boxes from them. However, the horizontal boxes are inclined to get small Intersection-over-Unions (IoUs) with ground truths, which may have some undesirable effects, such as introducing redundant noise, mismatching with ground truths, detracting from the robustness of detectors, etc. In this paper, we propose a novel Anchor-free Oriented Proposal Generator (AOPG) that abandons horizontal box-related operations from the network architecture. AOPG first produces coarse oriented boxes by a Coarse Location Module (CLM) in an anchor-free manner and then refines them into high-quality oriented proposals. After AOPG, we apply a Fast R-CNN head to produce the final detection results. Furthermore, the shortage of large-scale datasets is also a hindrance to the development of oriented object detection. To alleviate the data insufficiency, we release a new dataset on the basis of our DIOR dataset and name it DIOR-R. Massive experiments demonstrate the effectiveness of AOPG. Particularly, without bells and whistles, we achieve the accuracy of 64.41%, 75.24% and 96.22% mAP on the DIOR-R, DOTA and HRSC2016 datasets respectively. Code and models are available at https://github.com/jbwang1997/AOPG.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Gong Cheng (78 papers)
  2. Jiabao Wang (24 papers)
  3. Ke Li (723 papers)
  4. Xingxing Xie (5 papers)
  5. Chunbo Lang (3 papers)
  6. Yanqing Yao (1 paper)
  7. Junwei Han (87 papers)
Citations (251)

Summary

Anchor-free Oriented Proposal Generator for Object Detection

The paper "Anchor-free Oriented Proposal Generator for Object Detection" introduces an advanced object detection framework aimed at improving oriented object detection within remote sensing images. The conventional methodologies in object detection leverage horizontal boxes, which are pervasive in deriving the oriented boxes that ultimately exhibit lower Intersection-over-Unions (IoUs) with ground truths, affecting the overall robustness and efficacy of detectors. This paper proposes a novel Anchor-free Oriented Proposal Generator (AOPG) to bypass horizontal boxes entirely, thereby enhancing both proposal quality and detection accuracy through an innovative network architecture.

Key Contributions

  1. Innovative Anchor-free Proposal Generation: The AOPG method eliminates the horizontal anchor dependency prevalent in many existing object detectors. Traditional methods suffer from low IoUs and substantial shape inconsistencies introduced by horizontal anchors. AOPG circumvents this by directly proposing coarse oriented boxes via a Coarse Location Module (CLM), following an anchor-free approach. These initial proposals are subsequently refined into high-quality oriented proposals, enhancing alignment with true object orientations.
  2. Empirical Evaluation and Results: Profound experiments demonstrate marked improvements achieved by AOPG, as it realizes detection accuracies of 64.41%, 75.24%, and 96.22% on the DIOR-R, DOTA, and HRSC2016 datasets, respectively. These achievements highlight AOPG’s capability to outperform state-of-the-art object detection frameworks, especially in remote sensing, where object orientations are varied and complex.
  3. Development of a New Dataset - DIOR-R: Addressing the scarcity of suitable datasets for oriented object detection, the authors introduced DIOR-R, an enhancement over the existing DIOR dataset. This large-scale dataset substantially supports oriented object detection development, encompassing over 23,463 images and 192,518 instances across 20 categories.

Technical Implementation

The foundation of the proposed framework lies in a Feature Pyramid Network (FPN) backbone, extended by a specialized Coarse Location Module. This module derives coarse oriented boxes without presuming predefined anchors. By integrating the AlignConv operation, the network enhances alignment between features and irregularly-shaped objects, ensuring refined and high-quality oriented proposals. The final step involves applying a Fast R-CNN head to determine object classification and further regress oriented bounding boxes.

Implications and Future Directions

The anchor-free approach poses transformative implications for the structured methodology of object detection in AI environments. Moving beyond horizontal boxes aligns the detection process more closely with real-world complexities, particularly in domains such as surveillance and aerial imagery analysis. In the space of AI research and applications, this methodology could significantly enhance both the precision of object localization and classification accuracy.

Looking forward, potential advancements could include exploring the robustness of AOPG within varied environmental conditions or integrating it with other deep learning architectures. Furthermore, advancements might explore its application on other challenging domains, expanding our understanding of object detection in diverse settings.

This pivot away from traditional anchor methodologies proposes a compelling stride in oriented object detection, as shown through rigorous experimental validation. As research continues, evolving networks akin to AOPG could redefine the benchmarks in detecting and classifying arbitrary-oriented objects across numerous application areas within AI.

Github Logo Streamline Icon: https://streamlinehq.com