Source-linked AI summary
Strip R-CNN: Large Strip Convolution for Remote Sensing Object Detection
Xinbin Yuan, Zhaohui Zheng, Yuxuan Li, Xialei Liu, Li Liu, Xiang Li, Qibin Hou, Ming-Ming Cheng
TL;DR
Remote sensing object detection remains difficult for prevalent high-aspect-ratio objects, whose anisotropic features and geometry challenge existing detectors. Strip R-CNN uses orthogonal large strip convolutions in StripNet and a strip-enhanced localization head, achieving state-of-the-art performance across several benchmarks, including 82.75% mAP on DOTA-v1.0 with its 30M model.
Problem
High-aspect-ratio objects are prevalent in remote sensing, but existing detectors struggle with them as detection performance decreases with increasing aspect ratio.
Method
Strip R-CNN uses sequential orthogonal large strip convolutions in StripNet and strip convolutions in a decoupled localization branch.
Results
82.75% mAP is achieved on DOTA-v1.0 by the 30M Strip R-CNN-S model, while experiments on FAIR1M, HRSC2016, and DIOR show superiority over other methods.
Takeaways & Limitations
Large strip convolutions provide a simple, efficient approach for extracting features and improving localization across remote sensing objects with varied aspect ratios.
Abstract
from arXiv · showhide
While witnessed with rapid development, remote sensing object detection remains challenging for detecting high aspect ratio objects. This paper shows that large strip convolutions are good feature representation learners for remote sensing object detection and can detect objects of various aspect ratios well. Based on large strip convolutions, we build a new network architecture called Strip R-CNN, which is simple, efficient, and powerful. Unlike recent remote sensing object detectors that leverage large-kernel convolutions with square shapes, our Strip R-CNN takes advantage of sequential orthogonal large strip convolutions in our backbone network StripNet to capture spatial information. In addition, we improve the localization capability of remote-sensing object detectors by decoupling the detection heads and equipping the localization branch with strip convolutions in our strip head. Extensive experiments on several benchmarks, for example DOTA, FAIR1M, HRSC2016, and DIOR, show that our Strip R-CNN can greatly improve previous work. In particular, our 30M model achieves 82.75% mAP on DOTA-v1.0, setting a new state-of-the-art record. Our code will be made publicly available.Code is available at https://github.com/YXB-NKU/Strip-R-CNN.
1. Introduction
Remote sensing detectors still struggle with prevalent high-aspect-ratio objects because square-window features inadequately capture anisotropic context and their geometry complicates regression. Strip R-CNN addresses these challenges with orthogonal large strip convolutions and a strip-enhanced localization head.
- Motivation: High-aspect-ratio objects are common in remote sensing, yet existing detectors’ performance decreases as object aspect ratio increases.DOTA statistics show slender objects occupy a substantial portion of the data.
- Challenges: Square-window convolutions restrict anisotropic context capture by including irrelevant surrounding information around slender objects.Such objects contain rich information along one spatial dimension but sparse information along the other.
- Challenges: High-aspect-ratio objects also create substantial regression challenges because of their distinctive geometric properties.
- Approach: Strip R-CNN uses orthogonal large strip convolutions in StripNet to combine square and strip convolution advantages with less feature redundancy.The architecture is designed to remain simple and efficient while handling varied aspect ratios.
- Approach: The strip head decouples detection heads and strengthens localization with strip convolutions, improving localization and box-angle regression.
- Results: 82.75% mAP is achieved by the 30M-parameter Strip R-CNN-S on DOTA, establishing a new state-of-the-art record.The paper also reports experiments on FAIR1M, HRSC2016, and DIOR.
2. Related Work
Prior remote-sensing detectors use rotated boxes and increasingly large-kernel convolutions, but high-aspect-ratio objects remain difficult to detect. Strip R-CNN addresses this gap with sequential strip convolutions, a simpler StripNet backbone, and a strip-enhanced detection head.
- Rotated bounding boxes are widely used for arbitrarily oriented objects, but their angle periodicity can cause boundary discontinuity during regression.
- High-aspect-ratio objects are prevalent in remote-sensing datasets, yet existing detectors often lose performance as object slenderness increases.
- Large-kernel methods capture long-range context, but parallel large-kernel designs can increase computational burden and architectural complexity.
- StripNet backbone: Strip R-CNN uses sequential standard and orthogonal strip convolutions to capture features across objects with different aspect ratios without an additional fusion module.Its StripNet backbone replaces the O-RCNN backbone and uses residual blocks containing strip and feed-forward subblocks.
- Strip head: The strip head decouples classification, angle prediction, and localization, while the localization branch incorporates strip convolutions to capture long-range spatial correlations.Classification uses fully connected layers, whereas the localization branch is strengthened with the strip module.
4. Experiments
Across five remote sensing benchmarks, Strip R-CNN consistently improves detection performance while using strip-convolution designs to represent objects across aspect ratios. Ablations show that large kernels, sequential orthogonal strips, and the strip head each contribute to performance and localization.
- Main Results: Strip R-CNN-S improves over PKINet-S by 1.67% on DOTA-v1.0 while using 0.4% fewer parameters and 74.3% of its computations.It also improves over LSKNet-S by 2.57% with fewer parameters and computations.
- Main Results: 82.75% mAP is achieved on DOTA-v1.0 by ensembling RTMDet and Strip R-CNN, setting a new state-of-the-art record.The single model reaches 82.28% mAP with multi-scale training and testing.
- Main Results: Strip R-CNN improves over prior methods across DOTA-v1.5, FAIR1M-v1.0, DIOR-R, and HRSC2016.Reported gains include 0.8% on DOTA-v1.5, 48.26% mAP on FAIR1M-v1.0, 2.80% over LSKNet on DIOR-R, and 98.70% mAP under the VOC2012 metric on HRSC2016.
- Ablation Studies: Larger kernels in shallow StripNet layers combined with smaller kernels in deeper layers produce good results, whereas smaller shallow-layer kernels cause significant drops.The feature visualizations show more precise localization information and sharper object boundaries for slender objects.
- Ablation Studies: Kernel size 19 provides the best strip-module result and yields more accurate features for high aspect ratio objects.Smaller kernels are ineffective, kernel size 15 is satisfactory, and further improvement is observed at 19.
- Ablation Studies: Sequential horizontal and vertical strip convolutions outperform parallel arrangements, while square convolution remains important for square-shaped objects.Removing depthwise square convolution causes a large performance drop, and the parallel arrangement lacks effective two-dimensional modeling.
- Ablation Studies: StripNet improves RoI Transformer by 7.11%, S2ANet by 0.9%, and R3Det by 2.54%, while also boosting YOLO and DETR frameworks.The results demonstrate compatibility across multiple remote sensing detection frameworks.
- Ablation Studies: The strip head improves other detectors, including a 5.0 mAP gain for ROI Transformer and additional gains for Rotated Faster R-CNN.The authors report that the strip head consistently improves performance across detectors.
5. Conclusions
Strip R-CNN uses large strip convolutions to address slender-object detection and reports strong generalization across remote sensing benchmarks. Its visualizations show activations for high aspect ratio objects.
- Strip R-CNN uses large strip convolutions to extract features and improve localization of slender remote-sensing objects.
- The method achieves state-of-the-art performance on several remote sensing benchmarks.
- Eigen-CAM visualizations show strong activations for high aspect ratio objects, supporting the reported effectiveness.
6. Dataset details
The evaluation uses multiple remote sensing datasets spanning varied scales, orientations, shapes, object categories, and annotation settings. DOTA-v1.5 adds extremely small instances and a new category relative to DOTA-v1.0.
- DOTA-v1.0 contains 2,806 images, 188,282 instances, and 15 categories with varied scales, orientations, and shapes.
- DOTA-v1.5 retains DOTA-v1.0’s images and splits while adding instances smaller than 10 pixels and the Container Crane category.
- FAIR1M-v1.0 contains 15,266 high-resolution images and more than 1 million instances annotated with oriented bounding boxes.
- HRSC2016 contains 1,061 aerial images for ship detection, split into 436 training, 181 validation, and 444 test images.
- DIOR-R contains 23,463 images with oriented bounding-box annotations and 192,518 total annotations.
7. Implementation details
The implementation applies multi-scale training and testing to several datasets by rescaling images and cropping them into overlapping patches. Exponential Moving Average is not used during pre-training or fine-tuning.
- Images are rescaled to 0.5, 1.0, and 1.5 scales, then cropped into 1024×1024 patches with 500-pixel overlap.
- The implementation does not use Exponential Moving Average during pre-training or fine-tuning.
8. FAIR1M benchmark results
The FAIR1M-v1.0 section presents fine-grained category comparisons with state-of-the-art methods. The supplied passage identifies the comparison table but does not report its values.
- Fine-grained category results on FAIR1M-v1.0 are compared with state-of-the-art methods in Table 14.
9. DOTA benchmark results
This section presents fine-grained category comparisons with state-of-the-art methods on DOTA-v1.0 and DOTA-v1.5.
- Fine-grained category results are compared with state-of-the-art methods on DOTA-v1.0.
- Fine-grained category results are also compared with state-of-the-art methods on DOTA-v1.5.
- The reported comparisons are provided in Tables 16 and an additional referenced table.
10. More results of the effectiveness of the strip head
The strip head and StripNet-S backbone improve multiple remote sensing detectors across frameworks, while strip convolutions target limitations of square large-kernel designs.
- Strip head effectiveness: 0.56% improvement for LSKNet and 0.3% for Oriented R-CNN show that the strip head improves both detectors.The comparisons use DOTA-v1.0 with single-scale training and testing.
- StripNet-S backbone effectiveness: 7.11% improvement for RoI Transformer, 0.88% for Oriented R-CNN, 0.9% for S2ANet, and 2.54% for R3Det support StripNet-S compatibility across frameworks.The evaluated frameworks include two-stage and one-stage remote sensing detectors.
- Comparison with large-kernel backbones: Large-kernel comparison methods can suffer from neglected information, feature confusion, or feature redundancy in remote sensing scenarios.The passage contrasts dilated or large-kernel approaches with the proposed method’s focus on remote sensing data and high-aspect-ratio objects.
LSKNet
Additional detection results show that Strip R-CNN handles high-aspect-ratio objects that several comparison methods often miss or misdetect.
- Detection results: Strip R-CNN-S successfully captures high-aspect-ratio objects in the reported detection results.
- Backbone evaluation: Table 17 evaluates the StripNet-S backbone on other remote sensing object-detection frameworks using ImageNet pretraining for 100 epochs.
- Detection results: Oriented R-CNN, LSKNet, and PKINet frequently miss or incorrectly detect high-aspect-ratio objects, whereas Strip R-CNN detects them effectively.
- Head evaluation: Table 18 evaluates the strip head on other remote sensing object detectors.