Poly Kernel Inception Networks for Remote Sensing Detection

Introduction 👋

This repository contains the implementations of PKINet (CVPR2024) and PKINet-v2 (ECCV2026) for remote sensing object detection.

• PKINet is the original poly-kernel backbone for remote sensing detection.
• PKINet-v2 further improves both detection accuracy and inference efficiency over PKINet.

Pipeline of PKINet.

Pipeline of PKINet-v2.

---

Papers 📄

PKINet (CVPR 2024)

• Poly Kernel Inception Network for Remote Sensing Detection: arXiv

PKINet-v2 (ECCV 2026)

• PKINet-v2: Towards Powerful and Efficient Poly-Kernel Remote Sensing Object Detection: arXiv

---

Model Zoo 📦

PKINet-v2

Pretrained Models

• PKINet-v2-T backbone(ImageNet, 300 epochs): Download
• PKINet-v2-S backbone(ImageNet, 300 epochs): Download

DOTA-v1.0

Model	Detector	mAP	Angle	Config	Dev-Config	Weights
PKINet-v2-S-Sqrt5	Oriented R-CNN	80.75	le90	config	develop	model

The following models are the standard PKINet-v2 model zoo results.

Model	Detector	mAP	Angle	Config	Dev-Config	Weights
PKINet-v2-T	Oriented R-CNN	79.36	le90	config	-	model
PKINet-v2-S	Oriented R-CNN	80.46	le90	config	develop	model
PKINet-v2-S	Rotated Faster R-CNN	79.02	le90	config	develop	model
PKINet-v2-S	Rotated FCOS	77.13	le90	config	develop	model
PKINet-v2-S	R3Det	77.64	oc	config	develop	model
PKINet-v2-S	RoI Transformer	79.87	le90	config	develop	model
PKINet-v2-S	S2ANet	79.68	le135	config	develop	model

DOTA-v1.5

Model	Detector	mAP	Angle	Config	Weights
PKINet-v2-S	Oriented R-CNN	73.57	le90	config	model

---

PKINet

Pretrained Models

• PKINet-T backbone (ImageNet, 300 epochs): Download
• PKINet-S backbone (ImageNet, 300 epochs): Download

DOTA-v1.0

Model	mAP	Angle	Config	Weights
PKINet-T (1024,1024,200)	77.87	le90	pkinet-t_fpn_o-rcnn_dotav1-ss_le90	model
PKINet-S (1024,1024,200)	78.39	le90	pkinet-s_fpn_o-rcnn_dotav1-ss_le90	model

DOTA-v1.5

Model	mAP	Angle	Config	Weights
PKINet-S (1024,1024,200)	71.47	le90	pkinet-s_fpn_o-rcnn_dotav15-ss_le90	model

---

Setup 🛠️

MMRotate-PKINet depends on PyTorch, MMCV, MMDetection, and MMEngine.

For more details, please refer to the MMRotate installation guide.

conda create --name pkinet python=3.8 -y
conda activate pkinet
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
pip install yapf==0.40.1
pip install -U openmim
mim install mmcv-full
mim install mmdet
mim install mmengine
git clone https://github.com/NUST-Machine-Intelligence-Laboratory/PKINet.git
cd PKINet
mim install -v -e .

---

Training and Testing 🚀

Please see get_started.md for the basic usage of MMRotate.

We also provide tutorials for:

• learn the basics
• learn the config
• customize dataset
• customize model
• useful tools

PKINet-v2 (taking pkinet-v2-s_fpn_o-rcnn as an example)

1. Download ImageNet pretrained backbone and place it into pretrain/

• PKINet-v2-S backbone(ImageNet, 300 epochs): Download

2. Train PKINet-v2

bash ./tools/dist_train.sh configs/pkinet-v2/pkinet-v2-s_fpn_o-rcnn_3x_dotav1-ss_le90.py 4

3. Convert a trained PKINet-v2 checkpoint to the deploy version

python tools/pkinet/convert_pkinet_to_deploy.py \
  configs/pkinet-v2/pkinet-v2-s_fpn_o-rcnn_3x_dotav1-ss_le90.py \
  work_dirs/pkinet-v2-s_fpn_o-rcnn_3x_dotav1-ss_le90/latest.pth \
  work_dirs/pkinet-v2-s_fpn_o-rcnn_3x_dotav1-ss_le90/pkinet-v2-s-deploy.pth

4. Test the deploy model

bash ./tools/dist_test.sh \
  configs/pkinet-v2/pkinet-v2-s_fpn_o-rcnn_3x_dotav1-ss_le90_deploy.py \
  work_dirs/pkinet-v2-s_fpn_o-rcnn_3x_dotav1-ss_le90/pkinet-v2-s-deploy.pth \
  4 \
  --format-only \
  --eval-options submission_dir=work_dirs/results/pkinet-v2-dota

5. Submit the result to DOTA-v1.0 server

Important

If you care more about inference speed, we strongly recommend PyTorch 2.4.1 + CUDA 11.8, which is more efficient for large-kernel inference for inference. However, the training stability of this version has not been thoroughly verified, so we do not recommend using it for training.

conda create --name pkinet-dev python=3.10 -y
conda activate pkinet-dev
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
pip install yapf==0.40.1
pip install -U openmim
mim install mmcv-full
mim install mmdet
mim install mmengine
git clone https://github.com/NUST-Machine-Intelligence-Laboratory/PKINet.git
cd PKINet
mim install -v -e .

PKINet

1. Train PKINet

bash ./tools/dist_train.sh configs/pkinet/pkinet-s_fpn_o-rcnn_dotav1-ss_le90.py 4

---

License 📜

This project is released under the Apache 2.0 license.

---

Citation 📚

If you find this repository useful, please consider citing:

@inproceedings{cai2024poly,
    author    = {Cai, Xinhao and Lai, Qiuxia and Wang, Yuwei and Wang, Wenguan and Sun, Zeren and Yao, Yazhou},
    title     = {Poly Kernel Inception Network for Remote Sensing Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {27706-27716}
}

@article{cai2026pkinetv2,
  title={PKINet-v2: Towards Powerful and Efficient Poly-Kernel Remote Sensing Object Detection},
  author={Cai, Xinhao and Liulei Li and Gensheng Pei and Zeren Sun and Yazhou Yao and Wenguan Wang},
  journal={arXiv preprint arXiv:2603.16341},
  year={2026}
}