PKINet-v2: Towards Powerful and Efficient Poly-Kernel Remote Sensing Object Detection
PKINet-v2 combines anisotropic strip and square kernels to build multi-scale, geometry-adaptive features, achieving 80.46% mAP and 54.6 FPS on DOTA-v1.0.
Key Findings
Methodology
PKINet-v2 employs a hybrid approach integrating anisotropic strip convolutions with isotropic square kernels to form a multi-scope receptive field. This design captures targets with diverse geometries and scales. The model introduces a Heterogeneous Kernel Re-parameterization (HKR) strategy, merging multi-branch training structures into a single depth-wise convolution for inference, greatly improving efficiency. Extensive experiments on DOTA-v1.0, v1.5, HRSC2016, and DIOR-R demonstrate state-of-the-art accuracy and a 3.9× inference speedup, validating the approach’s effectiveness.
Key Results
- On DOTA-v1.0, PKINet-v2 achieves 80.46% mAP, surpassing PKINet-v1 by 2.07%, with an inference speed of 54.6 FPS, a 3.9× acceleration.
- In DOTA-v1.5 and HRSC2016, it outperforms existing backbones, especially on small and elongated targets, confirming its geometric and scale adaptability.
- Ablation studies show that integrating anisotropic strip convolutions and HKR significantly boosts both detection accuracy and inference speed, with minimal accuracy loss.
Significance
This work addresses the longstanding challenge of detecting targets with diverse geometries and scales in remote sensing images. By unifying multi-scale and geometry-aware features, PKINet-v2 enables real-time, high-precision detection, crucial for urban planning, disaster response, and environmental monitoring. Its innovative architecture sets a new benchmark, bridging the gap between accuracy and efficiency, and paving the way for practical deployment in resource-constrained scenarios.
Technical Contribution
The key technical contributions include the novel fusion of anisotropic strip and isotropic square convolutions to form a flexible, multi-scale receptive field, and the HKR strategy that consolidates multi-branch models into a single efficient depth-wise convolution. This design maintains detection robustness across diverse target geometries while significantly reducing inference latency. The architecture also introduces hierarchical multi-scale feature aggregation, enhancing local detail preservation and global context understanding, which are critical for complex remote sensing scenes.
Novelty
PKINet-v2 is the first to integrate anisotropic strip convolutions with isotropic square kernels within a unified framework for remote sensing object detection. Its innovative HKR strategy effectively merges multi-branch training structures into a single, hardware-friendly convolution, addressing the speed bottleneck of previous multi-scale models. This dual innovation advances the state-of-the-art in both geometric adaptability and inference efficiency, representing a significant leap forward.
Limitations
- Despite improvements, the model may still struggle with extremely small targets or highly cluttered backgrounds, requiring further feature refinement.
- The multi-scale fusion, while effective, can introduce false positives in complex scenes, necessitating more sophisticated post-processing.
- HKR’s hardware compatibility might vary across platforms, and further optimization is needed for deployment on edge devices.
Future Work
Future directions include enhancing robustness against extreme scale and geometric variations, integrating self-supervised learning for better generalization, and optimizing the HKR strategy for diverse hardware platforms. Additionally, exploring multi-modal data fusion and unsupervised domain adaptation could extend PKINet-v2's applicability to broader remote sensing tasks.
AI Executive Summary
Remote sensing image analysis faces persistent challenges due to the extreme variability in target geometry and scale. Traditional backbone networks often struggle to simultaneously capture slender, elongated structures and large, regular-shaped objects, especially under real-time constraints. Existing solutions tend to focus on either anisotropic kernels for slender targets or large isotropic kernels for broader context, but rarely both in a unified framework. This gap hampers detection accuracy and inference efficiency, limiting practical deployment in urban monitoring, disaster management, and environmental surveillance.
To address these issues, PKINet-v2 introduces a novel backbone architecture that synergistically combines anisotropic strip convolutions with isotropic square kernels. This hybrid design constructs a multi-scope receptive field capable of modeling diverse geometries and scales. The model further employs a hierarchical feature aggregation strategy, ensuring the preservation of fine details while capturing global context. Crucially, the introduction of the Heterogeneous Kernel Re-parameterization (HKR) strategy transforms the complex multi-branch training architecture into a single, efficient depth-wise convolution during inference, significantly reducing computational overhead.
Experimental results on four widely-used remote sensing benchmarks—DOTA-v1.0, v1.5, HRSC2016, and DIOR-R—demonstrate that PKINet-v2 achieves state-of-the-art detection accuracy, with a mean Average Precision (mAP) of 80.46% on DOTA-v1.0, surpassing previous backbones. Meanwhile, inference speed is boosted by 3.9 times, reaching 54.6 FPS, enabling real-time applications. The ablation studies confirm that the combination of multi-scale, geometry-aware features and the HKR strategy are key to these improvements.
This work marks a significant step forward in remote sensing object detection, offering a model that is both highly accurate and computationally efficient. Its ability to adapt to complex target geometries and scales opens new possibilities for practical deployment in resource-constrained environments, such as onboard satellite or drone systems. Looking ahead, further research will focus on robustness enhancements, hardware optimization, and multi-modal data integration, aiming to push the boundaries of real-time, high-precision remote sensing analysis.
Deep Dive
Abstract
Object detection in remote sensing images (RSIs) is challenged by the coexistence of geometric and spatial complexity: targets may appear with diverse aspect ratios, while spanning a wide range of object sizes under varied contexts. Existing RSI backbones address the two challenges separately, either by adopting anisotropic strip kernels to model slender targets or by using isotropic large kernels to capture broader context. However, such isolated treatments lead to complementary drawbacks: the strip-only design can disrupt spatial coherence for regular-shaped objects and weaken tiny details, whereas isotropic large kernels often introduce severe background noise and geometric mismatch for slender structures. In this paper, we extend PKINet, and present a powerful and efficient backbone that jointly handles both challenges within a unified paradigm named Poly Kernel Inception Network v2 (PKINet-v2). PKINet-v2 synergizes anisotropic axial-strip convolutions with isotropic square kernels and builds a multi-scope receptive field, preserving fine-grained local textures while progressively aggregating long-range context across scales. To enable efficient deployment, we further introduce a Heterogeneous Kernel Re-parameterization (HKR) Strategy that fuses all heterogeneous branches into a single depth-wise convolution for inference, eliminating fragmented kernel launches without accuracy loss. Extensive experiments on four widely-used benchmarks, including DOTA-v1.0, DOTA-v1.5, HRSC2016, and DIOR-R, demonstrate that PKINet-v2 achieves state-of-the-art accuracy while delivering a $\textbf{3.9}\times$ FPS acceleration compared to PKINet-v1, surpassing previous remote sensing backbones in both effectiveness and efficiency.