Research map / Detection and segmentation / CNN features
Poly Kernel Inception Network for Remote Sensing Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PKINet, a lightweight backbone using dilation-free multi-scale depth-wise inception kernels plus a Context Anchor Attention module, achieving state-of-the-art oriented object detection on remote sensing benchmarks.
Problem
Remote sensing object detection struggles with large object scale variation and diverse-ranging context; large-kernel convolutions add background noise for small objects while dilated convolutions produce overly sparse features.
Approach
Replace large-kernel or dilated convolutions with parallel dilation-free depth-wise inception kernels (3x3 to 11x11) for multi-scale local context, and add a Context Anchor Attention module that uses average pooling and 1D strip depth-wise convolutions to capture long-range context.
Key results
- PKINet-S with Oriented RCNN reaches 78.39% mAP on DOTA-v1.0, versus 77.49% for LSKNet-S and 75.87% for ResNet-50 based Oriented RCNN.
- PKINet-T achieves 77.87% mAP on DOTA-v1.0 with 4.1M backbone parameters, versus 74.20% for ResNet-18 at 11.2M.
- PKINet-S reaches 71.47% mAP on DOTA-v1.5 (vs 70.26% LSKNet-S) and 98.54% on HRSC2016 (07 metric 90.70%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ResNet-50 with PKINet-S | Oriented RCNN on DOTA-v1.0 · DOTA-v1.0 | mAP: 75.87 -> 78.39 | Helped |
| replace LSKNet-S with PKINet-S | Oriented RCNN on DOTA-v1.5 · DOTA-v1.5 | mAP: 70.26 -> 71.47 | Helped |
| replace LSKNet with PKINet-S | Oriented RCNN on HRSC2016 · HRSC2016 | mAP: 90.65 -> 90.70; 98.46 -> 98.54 | Helped |
| replace DCFL with PKINet-S | Oriented RCNN on DIOR-R · DIOR-R | mAP: 66.80 -> 67.03 | Helped |
| replace ConvNeXt-N with PKINet-S | Mask R-CNN on COCO 2017 · COCO 2017 | mAP: 41.3 -> 43.4 | Helped |
| use multi-scale kernels instead of only 3×3 kernels (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 76.94 -> 78.16 | Helped |
| use five kernels instead of two, three, four, or six (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 75.76; 76.07; 77.53; 78.16; 78.10 | Helped |
| apply CAA in all stages rather than no stages or one stage (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 77.13 -> 78.16 | Helped |
| use dilation in the PKI kernels (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 78.16 -> 77.07; 78.16 -> 76.95 | Hurt |
| remove the CSP structure (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 78.16; 77.83 | Mixed |
| increase CAA kernel sizes, using the expansive kernel-size strategy (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 77.52; 77.71; 77.76; 77.89; 78.16 | Helped |
| use only large kernels instead of multi-scale kernels (ablation) | PKINet on DOTA-v1.0 · DOTA-v1.0 | mAP: 78.16 -> 77.67; 78.16 -> 77.32 | Hurt |
Surprising findings
- Dilated kernels reduced accuracy (about -1 point with dilation 2) despite a larger receptive field.
- Using only large kernels (11x11 across all blocks) performed worse than the multi-scale design, supporting the claim that large kernels add noise.
Limitations
- Model scalability was not explored due to limited computational resources.
- On HRSC2016 the gain over LSKNet is slight because the ship classes are merged into one category, which does not test scale variation well.
Experimental setup
Datasets: ImageNet-1K DOTA-v1.0 DOTA-v1.5 HRSC2016 DIOR-R COCO 2017
Benchmarks: DOTA-v1.0 DOTA-v1.5 HRSC2016 DIOR-R COCO 2017
Metrics: mAP AP50 parameters FLOPs
Baselines: ResNet-50 ResNet-18 LSKNet-S ARC Oriented RCNN ReDet RoI Transformer R3Det S2ANet Rotated FCOS
Builds on: Oriented RCNN LSKNet ConvNeXt CSP
Closest papers by meaning
- SM3Det: A Unified Model for Multi-Modal Remote Sensing Object Detection (AAAI 2026 · similarity 0.87)
- LWGANet: Addressing Spatial and Channel Redundancy in Remote Sensing Visual Tasks with Light-Weight Grouped Attention (AAAI 2026 · similarity 0.86)
- Strip R-CNN: Large Strip Convolution for Remote Sensing Object Detection (AAAI 2026 · similarity 0.85)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 2025 · similarity 0.83)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.82)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.