[Paper Review] SCRDet: Towards More Robust Detection for Small, Cluttered and Rotated Objects

Xue Yang, Jirui Yang+6 more|arXiv (Cornell University)|Nov 17, 2018

Advanced Neural Network ApplicationsComputer Science47 references145 citations

TL;DR

SCRDet presents a multi-category rotation detector for small, cluttered, and arbitrarily oriented objects, introducing SF-Net for finer sampling, MDA-Net for supervised attention, and an IoU-enhanced rotation loss to improve rotated object detection. It achieves state-of-the-art results on remote sensing and general datasets.

ABSTRACT

Object detection has been a building block in computer vision. Though considerable progress has been made, there still exist challenges for objects with small size, arbitrary direction, and dense distribution. Apart from natural images, such issues are especially pronounced for aerial images of great importance. This paper presents a novel multi-category rotation detector for small, cluttered and rotated objects, namely SCRDet. Specifically, a sampling fusion network is devised which fuses multi-layer feature with effective anchor sampling, to improve the sensitivity to small objects. Meanwhile, the supervised pixel attention network and the channel attention network are jointly explored for small and cluttered object detection by suppressing the noise and highlighting the objects feature. For more accurate rotation estimation, the IoU constant factor is added to the smooth L1 loss to address the boundary problem for the rotating bounding box. Extensive experiments on two remote sensing public datasets DOTA, NWPU VHR-10 as well as natural image datasets COCO, VOC2007 and scene text data ICDAR2015 show the state-of-the-art performance of our detector. The code and models will be available at https://github.com/DetectionTeamUCAS.

Motivation & Objective

• Motivate robust detection of small, cluttered, and arbitrarily oriented objects in aerial and natural images.
• Develop a detector that combines tailored sampling, attention to suppress background noise, and rotation-aware regression.
• Demonstrate that the proposed techniques generalize across remote sensing and natural image datasets.
• Achieve state-of-the-art performance on public benchmarks (DOTA, NWPU VHR-10) and competitive results on COCO, VOC2007, ICDAR2015.

Proposed method

• Propose SF-Net to enable finer anchor sampling and feature fusion for small objects by using smaller anchor stride and multi-layer feature fusion.
• Introduce a supervised multi-dimensional attention network (MDA-Net) with pixel and channel attention to suppress background noise and highlight foreground objects.
• Add a rotation-aware branch with five-parameter (x, y, w, h, theta) regression and rotation non-maximum suppression (R-NMS) using skew IoU for accurate oriented bounding boxes.
• Modify the regression loss by incorporating an IoU-based factor into the smooth L1 loss to address boundary discontinuities in rotated boxes.
• Train with a multi-task loss combining rotated box regression, attention supervision, and classification losses.
• Demonstrate generality by validating on both remote sensing (DOTA, NWPU VHR-10) and natural image datasets (COCO, VOC2007, ICDAR2015).

Experimental results

Research questions

• RQ1 How can small objects be more effectively sampled and localized in multi-category remote sensing scenes?
• RQ2 Can a supervised attention mechanism improve detection performance in cluttered, noisy backgrounds?
• RQ3 Does an IoU-enhanced rotation loss stabilize and improve regression for arbitrarily oriented bounding boxes?
• RQ4 Do the proposed components generalize to natural image datasets beyond remote sensing?
• RQ5 What is the overall performance gain of SCRDet on standard benchmarks for oriented and horizontal bounding boxes?

Key findings

• SCRDet achieves state-of-the-art performance on OBB detection for DOTA with a mAP of 72.61% on the proposed configuration.
• SCRDet achieves the best performance among published methods on NWPU VHR-10 for horizontal bounding box detection with a mAP of 91.75%.
• In ablations, MDA-Net provides notable gains (and up to around 3.7 percentage points in mAP on DOTA) by suppressing noise and highlighting object cues.
• SF-Net with finer sampling and feature fusion yields strong improvements for small objects, achieving the best overall mAP (68.89% in one ablation).
• IoU-smooth L1 loss addresses rotation boundary discontinuities and improves detection accuracy (e.g., up to 69.83% mAP in ablations).
• On natural image datasets, SCRDet-augmented baselines (e.g., R2CNN) achieve higher single-scale mAPs (e.g., 80.08% on ICDAR2015) demonstrating generality of the approach.