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

Xinhao Cai; Liulei Li; Gensheng Pei; Zeren Sun; Yazhou Yao; Wenguan Wang

cs.CVMar 17, 2026

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

Authors: Xinhao Cai, Liulei Li, Gensheng Pei, Zeren Sun, Yazhou Yao, Wenguan Wang

Organizations: Nanjing University of Science and Technology · State Key Lab. of Intelligent Manufacturing of Advanced Construction Machinery · Zhejiang University · Department of Electrical and Computer Engineering, Sungkyunkwan University

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.

Figures & tables

Panel of Figure 1 — Figure 1: Performance of PKINet-v2 on DOTA-v1 dataset [ 55 ] . PKINet-v2 consistently improves mAP while substantially boosting inference efficiency over PKINet-v1 [ 1 ] , across representative oriented detectors.

Backbone	mAP $\uparrow$	FPS $\uparrow$	# P $\downarrow$	FLOPs $\downarrow$
LSKNet-S [ 30 ]	77.49	51.00	31.0M	161G
StripNet-S [ 68 ]	79.85	46.40	30.5M	172G
LWGANet-L2 [ 38 ]	79.02	33.50	29.2M	159G
PKINet-v1-S [ 1 ]	78.39	14.05	30.8M	184G
PKINet-v2-S	80.46	54.60	30.7M	173G

Table 1: Remote Sensing Backbones Comparison . All backbones are pretrained on ImageNet-1K [ 8 ] for 300 epochs and built within the framework of Oriented RCNN [ 56 ] . FPS is tested on a single NVIDIA A100-40G GPU.

Figure 2: PKINet-v2 overview. PKINet-v2 consists of four stages. Each (a) Stage $l$ contains a Patch Embedding layer followed by a sequence of $N_{l}$ PKINet-v2 Blocks. Each (b) PKINet-v2 Block is composed of a PKS sub-block and an FFN sub-block. The (c) PKS Block mainly comprises a PKS Module and two Fully Connected (FC) layers. (d) PKS Module (§ 3.3 ) builds a multi-scope spatial attention map by aggregating heterogeneous branches with wide-range receptive field and fusing them with a $1\times 1$ convolution. Here, $n\!=\!0,\dots,N_{l}\!-\!1$ indicates that the PKS Module/Block is located in the $n$ -th PKINet-v2 Block of the $l$ -th stage. Please refer to § 3 for more details.

Figure 3: Receptive field construction and HKR Strategy. (Left) The multi-branch PKS Module (§ 3.3 ) aggregates heterogeneous depth-wise branches to form a hierarchically densified receptive field with a broad receptive-field range, where global full-span coverage is progressively enriched toward the center. (Right) Heterogeneous Kernel Re-parameterization (HKR, § 3.4 ) algebraically fuses Conv-BN and merges all branches into one equivalent $K_{\max}\!\times\!K_{\max}$ depth-wise convolution, preserving identical outputs while significantly improving inference efficiency. Please refer to § 3 for details.

Variant	{ $C_{1}$ , $C_{2}$ , $C_{3}$ , $C_{4}$ }	{ $D_{1}$ , $D_{2}$ , $D_{3}$ , $D_{4}$ }	#P	FLOPs	FPS	FPS-HKR
$\star$ PKINet-v2-T	{32, 64, 160, 256}	{3, 3, 5, 2}	4.0M	18.2G	46.2	58.0
$\star$ PKINet-v2-S	{64, 128, 320, 512}	{2, 2, 4, 2}	13.6M	54.0G	43.4	54.6

Table 2: Variants of PKINet-v2 backbone.

C_{i}

: feature channel number;

D_{i}

: number of strip blocks in each stage

i

. All metrics are reported with an input resolution of

1024\times 1024

, and FPS is tested on a single NVIDIA A100-40G GPU. All variants are built within the framework of Oriented RCNN [ 56 ] . Params and FLOPs are computed for backbones only. Please refer to § 3 for more details.

Method	mAP $\uparrow$	# P $\downarrow$	FLOPs $\downarrow$	PL	BD	BR	GTF	SV	LV	SH	TC	BC	ST	SBF	RA	HA	SP	HC
DETR-based
AO 2 -DETR [ 7 ]	70.91	40.8M	304G	87.99	79.46	45.74	66.64	78.90	73.90	73.30	90.40	80.55	85.89	55.19	63.62	51.83	70.15	60.04
O 2 -DETR [ 41 ]	72.15	-	-	86.01	75.92	46.02	66.65	79.70	79.93	89.17	90.44	81.19	76.00	56.91	62.45	64.22	65.80	58.96
ARS-DETR [ 69 ]	73.79	41.6M		86.61	77.26	48.84	66.76	78.38	78.96	87.40	90.61	82.76	82.19	54.02	62.61	72.64	72.80	64.96
One-stage
SASM [ 23 ]	74.92	36.6M	194G	86.42	78.97	52.47	69.84	77.30	75.99	86.72	90.89	82.63	85.66	60.13	68.25	73.98	72.22	62.37

Table 3: Experimental results on DOTA-v1.0 dataset [ 55 ] under single-scale training and testing setting. PKINet-v2-S is pretrained on ImageNet-1K [ 8 ] for 300 epochs similar to previous methods [ 30 , 1 , 68 , 38 ] . † : Use the Strip Head [ 68 ] . See § 4.2 .

Method	mAP $\uparrow$	PL	BD	BR	GTF	SV	LV	SH	TC	BC	ST	SBF	RA	HA	SP	HC	CC
RetinaNet-O [ 32 ]	59.16	71.43	77.64	42.12	64.65	44.53	56.79	73.31	90.84	76.02	59.96	46.95	69.24	59.65	64.52	48.06	0.83
FR-O [ 48 ]	62.00	71.89	74.47	44.45	59.87	51.28	68.98	79.37	90.78	77.38	67.50	47.75	69.72	61.22	65.28	60.47	1.54
Mask RCNN [ 20 ]	62.67	76.84	73.51	49.90	57.80	51.31	71.34	79.75	90.46	74.21	66.07	46.21	70.61	63.07	64.46	57.81	9.42
HTC [ 2 ]	63.40	77.80	73.67	51.40	63.99	51.54	73.31	80.31	90.48	75.12	67.34	48.51	70.63	64.84	64.48	55.87	5.15
ReDet [ 19 ]	66.86	79.20	82.81	51.92	71.41	54.08	75.73	80.92	90.83	75.81	68.64	49.29	72.03	73.36	70.55	63.33	11.53
DCFL [ 57 ]	67.37	-	-	-	-	56.72	-	80.87	-	-	75.65	-	-	-	-	-	-

Table 4: Experimental results on DOTA-v1.5 dataset [ 55 ] compared with state-of-the-art methods with single-scale training and testing. PKINet-v2-S backbone is pretrained on ImageNet-1K [ 8 ] for 300 epochs, as the compared methods [ 19 , 57 , 1 , 68 ] . PKINet-v2-S is built within the framework of Oriented RCNN [ 56 ] . See § 4.2 for details.

Table 8

Method	Backbone	mAP $\uparrow$	#P $\downarrow$	FLOPs $\downarrow$
Rotated FCOS [ 51 ]	ResNet-50 [ 21 ]	72.45	41.1M	211.4G
	PKINet-v1 [ 1 ]	74.86	13.7M	70.2G
	PKINet-v2	77.13	13.6M	54.0G
R3Det [ 60 ]	ResNet-50 [ 21 ]	69.70	41.1M	211.4G
	ARC-R50 [ 45 ]	72.32	56.5M	86.6G
	PKINet-v1-S [ 1 ]	75.89	13.7M	70.2G

Table 7: Experimental results across different detection frameworks on DOTA-v1.0 dataset [ 55 ] . Refer to § 4.2 .

Table 8: Performance analysis on DOTA-v1.0 [ 55 ] . (Left) mAP across object aspect-ratio bins. (Right) mAP across object-size bins (relative area). All models are trained on the train set and evaluated on the val set. Please refer to § 4.2 for details.

Figure 4: Visual results on DOTA-1.0 dataset [ 55 ] . (Top) Geometric complexity. (Bottom) Spatial complexity. PKINet-v2 delivers more robust detection than other methods [ 1 , 68 , 30 ] under both challenges. Please refer to § 4.3 for more details.

Table 9: Diagnostic experiments on DOTA-v1.0 [ 55 ] . The adopted default designs are marked in red . All models are pretrained on ImageNet-1K [ 8 ] for 100 epochs (ablation setting) and built within Oriented RCNN [ 56 ] . FPS reported in Tab. 9(a) and Tab. 9(c) is measured during inference after applying HKR. See § 4.4 for details.

Method	Backbone	mAP $\uparrow$	# P $\downarrow$	PL	BD	BR	GTF	SV	LV	SH	TC	BC	ST	SBF	RA	HA	SP	HC
DETR-based
AO 2 -DETR [ 7 ]	ResNet-50 [ 21 ]	70.91	40.8M	87.99	79.46	45.74	66.64	78.90	73.90	73.30	90.40	80.55	85.89	55.19	63.62	51.83	70.15	60.04
O 2 -DETR [ 41 ]	ResNet-50 [ 21 ]	72.15	-	86.01	75.92	46.02	66.65	79.70	79.93	89.17	90.44	81.19	76.00	56.91	62.45	64.22	65.80	58.96
ARS-DETR [ 69 ]	ResNet-50 [ 21 ]	73.79	41.6M	86.61	77.26	48.84	66.76	78.38	78.96	87.40	90.61	82.76	82.19	54.02	62.61	72.64	72.80	64.96
One-stage
SASM [ 23 ]	ResNet-50 [ 21 ]	74.92	36.6M	86.42	78.97	52.47	69.84	77.30	75.99	86.72	90.89	82.63	85.66	60.13	68.25	73.98	72.22	62.37

Table S1 : Experimental results on DOTA-v1.0 dataset [ 55 ] across different detection frameworks under the single-scale training and testing setting. PKINet-v2 is pretrained on ImageNet-1K [ 8 ] for 300 epochs similar to previous methods [ 1 , 30 , 68 ] .

Figure S1 : More qualitative comparisons on DOTA [ 55 ] on Oriented R-CNN [ 56 ] with PKINet-v1 [ 1 ] . See § 4 for details.

Figure S2 : More qualitative results on DOTA [ 55 ] on Oriented R-CNN [ 56 ] of PKINet-v2. See § 4 for details.

Figure S3 : Representative failure cases on DOTA [ 55 ] over Oriented R-CNN [ 56 ] . See § 4 for details.

Explore similar work

RMK RetinaNet: Rotated Multi-Kernel RetinaNet for Robust Oriented Object Detection in Remote Sensing Imagery
Mar 5, 2026Huiran SunObject Detection Aerial Object Detection
School of Computer Science and Engineering, Changchun University of Technology, 2055 Yan’an Street, Chaoyang District, Changchun, 130012, Jilin Province, China
BMCR: Adaptive Backbone Module Composition via Reinforcement Learning for Remote Sensing Object Detection
Jun 4, 2026Wenlin Liu, Xikun Hu, Ping ZhongReinforcement Learning Dynamic Neural Networks
College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China
SPEANet: Structural Prior Enhanced Attention Network for Parameter-Efficient Remote Sensing Object Detection
Aug 15, 2026Wei Lu, Junjie Li, Feifei Sang +1Efficient Object Detection Aerial Object Detection
MOE Key Lab of ICSP, IMIS Lab of Anhui, Anhui Provincial Key Lab of Multimodal Cognitive Computation, Zenmorn-AHU · AI Joint Lab, School of Computer Science and Technology, Anhui University, Hefei 230601, China

Mar 5, 2026J/K move · Enter open · S save

RMK RetinaNet: Rotated Multi-Kernel RetinaNet for Robust Oriented Object Detection in Remote Sensing Imagery

Huiran Sun

School of Computer Science and Engineering, Changchun University of Technology, 2055 Yan’an Street, Chaoyang District, Changchun, 130012, Jilin Province, China

Object Detection Aerial Object Detection Multi-Scale Feature Fusion

Open paper AskPDF arXiv

Rotated object detection in remote sensing imagery is hindered by three major bottlenecks: non-adaptive receptive field utilization, inadequate long-range multi-scale feature fusion, and discontinuities in angle regression. To address these issues, we propose Rotated Multi-Kernel RetinaNet (RMK RetinaNet). First, we design a Multi-Scale Kernel (MSK) Block to strengthen adaptive multi-scale feature extraction. Second, we incorporate a Multi-Directional Contextual Anchor Attention (MDCAA) mechanism into the feature pyramid to enhance contextual modeling across scales and orientations. Third, we introduce a Bottom-up Path to preserve fine-grained spatial details that are often degraded during downsampling. Finally, we develop an Euler Angle Encoding Module (EAEM) to enable continuous and stable angle regression. Extensive experiments on DOTA-v1.0, HRSC2016, and UCAS-AOD show that RMK RetinaNet achieves performance comparable to state-of-the-art rotated object detectors while improving robustness in multi-scale and multi-orientation scenarios.

Abstract

Explore similar work

RMK RetinaNet: Rotated Multi-Kernel RetinaNet for Robust Oriented Object Detection in Remote Sensing Imagery

BMCR: Adaptive Backbone Module Composition via Reinforcement Learning for Remote Sensing Object Detection

SPEANet: Structural Prior Enhanced Attention Network for Parameter-Efficient Remote Sensing Object Detection