RiO-DETR: DETR for Real-time Oriented Object Detection

Zhangchi Hu; Yifan Zhao; Yansong Peng; Wenzhang Sun; Xiangchen Yin; Jie Chen; Peixi Wu; Hebei Li

cs.CVMar 10, 2026

RiO-DETR: DETR for Real-time Oriented Object Detection

Authors: Zhangchi Hu, Yifan Zhao, Yansong Peng, Wenzhang Sun, Xiangchen Yin, Jie Chen, Peixi Wu, Hebei Li, +3 more

Organizations: University of Science and Technology of China · Tsinghua University · Huawei Technologies Co., Ltd. · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center

Abstract

We present RiO-DETR: DETR for Real-time Oriented Object Detection, the first real-time oriented detection transformer to the best of our knowledge. Adapting DETR to oriented bounding boxes (OBBs) poses three challenges: semantics-dependent orientation, angle periodicity that breaks standard Euclidean refinement, and an enlarged search space that slows convergence. RiO-DETR resolves these issues with task-native designs while preserving real-time efficiency. First, we propose Content-Driven Angle Estimation by decoupling angle from positional queries, together with Rotation-Rectified Orthogonal Attention to capture complementary cues for reliable orientation. Second, Decoupled Periodic Refinement combines bounded coarse-to-fine updates with a Shortest-Path Periodic Loss for stable learning across angular seams. Third, Oriented Dense O2O injects angular diversity into dense supervision to speed up angle convergence at no extra cost. Extensive experiments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 demonstrate RiO-DETR establishes a new speed--accuracy trade-off for real-time oriented detection. GitHub Repository: https://github.com/RicePasteM/RiO-DETR.

Figures & tables

Figure 1: Comparisons with other detectors in terms of model size (left), latency (mid), and computational cost (right) on DOTA-1.0 under single-scale training and testing protocol. * denotes a community implemented version.

Figure 2: The main architecture of our proposed RiO-DETR. The framework highlights three key components: (A) Content-Driven Angle Estimation, (B) Rotation-Rectified Orthogonal Attention, and (C) Decoupled Periodic Refinement.

Figure 3: An intuitive illustration of (D) Decoupled Periodic Refinement and (E) Oriented Dense O2O.

Table 4

Table 5

Methods	#P	Flops	Latency	AP 50
Oriented RT-DETRv2 (Ours Scratch Implementation)	18.61M	97.06G	5.04	70.35
+ Hausdorff Matching from RHINO-DETR [ 24 ] & KLD Loss [ 55 ]				72.86
+ Universal Matching Strategy from D-FINE [ 35 ]				73.33
+ Dense O2O from DEIM [ 21 ] ( Our Implemented Baseline )	18.61M	97.06G	5.04	73.47
+ Content-Driven Angle Estimation
Geometry-Decoupled Query Encoding	18.59M	97.01G	5.04	74.18

Table 5: Step-by-step modifications from baseline model to RiO-DETR-m on DIOR-R. Each step shows changes in AP 50 , the number of parameters, latency, and Flops.

Table 7

Strategy	Ratio ( $\theta:\theta+\pi/2$ )	AP 50
Vanilla Alignment	$8:0$	73.81
Asymmetric Alignment	$6:2$	74.02
Multi-angle Distribution	$2:2:2:2^{\ast}$	73.95
Symmetric Orthogonal	$4:4$	74.18
∗ Multi-angle includes $\theta,\theta+\pi/4,\theta+\pi/2$ , and $\theta+3\pi/4$ .

Table 10: Ablation study on Attention Head Splitting Strategy for Rotation-Rectified Orthogonal Attention. H denotes the number of attention heads (default H = 8)

Figure 4: Visualization of deformable attention sampling points on DOTA.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	RiO-DETR-x	RiO-DETR-l	RiO-DETR-m	RiO-DETR-s	RiO-DETR-n
Backbone Name	HGNetv2-B5	HGNetv2-B4	HGNetv2-B2	HGNetv2-B0	HGNetv2-B0
Optimizer	AdamW	AdamW	AdamW	AdamW	AdamW
Embedding Dimension	384	256	256	224	128
Feedforward Dimension	2048	1024	1024	1024	512
Decoder Layers	4	4	3	3	3
Queries	300	300	300	300	300

Appendix

Table 1: Hyperparameter configurations for RiO-DETR models on DOTA-1.0.

Method	Backbone	AP 50	AP 75
Oriented R-CNN [ 49 ]	R-50	74.19	46.96
RoI Trans. [ 8 ]	R-50	74.05	46.54
RoI Trans. [ 8 ]	Swin-T	76.49	50.15
ReDet [ 16 ]	ReR-50	76.25	50.86
ARS-DETR [ 58 ]	R-50	73.79	49.01
ARS-DETR [ 58 ]	Swin-T	75.79	51.11

Appendix

Table 2: Comparison on DOTA-1.0 under single-scale training and testing protocol. Left: high-precision comparison. Right: results of RiO-DETR on other backbones.

Methods	Backbone	#P	Flops	Lat.	AP 50	PL	BD	BR	GTF	SV	LV	SH	TC	BC	ST	SBF	RA	HA	SP	HC
Non-Real-time Oriented Object Detectors
PKINet-S [ 1 ]	PKINet-S	30.8M	190G	187.5	81.06	89.0	86.7	59.0	81.2	80.4	84.9	88.1	90.9	86.6	87.3	67.1	74.8	78.2	81.9	70.6
LSKNet-S [ 27 ]	LSKNet-S	31.0M	161G	146.2	81.64	89.6	86.3	63.1	83.7	82.2	86.1	88.7	90.9	88.4	87.4	71.7	69.6	78.9	81.8	76.5
Strip R-CNN-S [ 57 ]	StripNet-S	30.5M	159G	106.3	82.28	89.2	85.6	62.4	83.7	81.9	86.6	88.8	90.9	88.0	87.9	72.1	71.9	79.2	82.5	82.8
Real-time Oriented Object Detectors
PPYOLOE-R-x [ 51 ]	CRN-x	98.4M	264G	32.8	80.73	88.5	84.5	60.6	77.7	83.3	85.4	89.0	90.8	88.5	87.5	69.3	66.0	77.9	81.4	80.9

Appendix

Table 3: Comparison with state-of-the-art oriented object detectors on DOTA-1.0 under multi-scale training and testing protocol.

Method	C1 C20	C2 C21	C3 C22	C4 C23	C5 C24	C6 C25	C7 C26	C8 C27	C9 C28	C10 C29	C11 C30	C12 C31	C13 C32	C14 C33	C15 C34	C16 C35	C17 C36	C18 C37	C19 –	AP 50 (%)
SASM RepPoints ∗	38.3 0.1	61.7 2.4	15.0 37.2	35.4 80.4	0.4 49.9	44.0 87.4	53.3 44.5	30.6 54.8	20.7 28.3	3.9 14.7	70.7 31.9	53.6 14.3	9.4 24.8	29.9 17.7	21.2 25.7	48.3 48.4	2.8 32.2	0.1 5.9	2.2 –	30.9
R-FCOS ∗	39.2 5.4	82.7 2.7	15.9 43.7	52.0 86.2	0.2 56.3	41.0 88.1	54.5 48.0	44.2 57.2	43.5 25.7	9.3 24.0	73.1 43.8	55.1 22.0	11.5 18.9	37.9 31.5	26.0 33.4	51.0 56.8	6.3 27.3	1.1 7.3	12.9 –	36.1
S2A-Net ∗	40.4 2.5	83.7 2.2	17.2 46.2	49.9 82.1	0.2 59.0	42.5 88.2	57.0 41.5	42.7 62.9	52.0 20.3	10.3 24.8	74.8 42.9	66.2 19.9	13.4 24.8	39.7 32.5	35.7 33.3	61.7 58.5	2.8 34.9	0.8 6.6	10.5 –	37.4
R-Faster R-CNN ∗	37.5 11.9	82.5 2.0	15.4 42.6	47.7 83.3	9.4 55.4	41.7 89.4	50.8 50.4	47.5 65.1	60.6 26.1	8.3 19.0	72.5 47.5	57.7 17.9	12.9 24.2	41.4 26.3	38.2 30.5	53.4 54.1	11.7 26.3	2.1 8.0	16.8 –	37.5
O-RepPoints ∗	37.7 5.7	81.4 0.9	15.4 46.3	48.8 84.3	7.9 61.4	43.4 89.2	54.6 46.7	43.5 60.8	56.6 33.2	9.5 24.6	73.3 50.4	62.9 24.9	15.4 22.2	42.0 32.8	42.8 34.7	59.5 60.0	4.5 37.9	0.6 7.4	17.0 –	38.9
O-RCNN ∗	37.6 10.4	84.8 1.2	16.9 47.2	49.0 83.3	7.3 59.6	41.6 89.1	49.7 48.3	43.1 62.6	62.4 31.8	13.1 26.6	72.1 58.5	59.8 28.2	18.5 25.5	43.9 38.8	41.5 37.3	56.0 64.6	7.9 38.7	2.2 10.7	24.4 –	40.4

Appendix

Table 4: Comparison with state-of-the-art models on the FAIR-1M-2.0 dataset. The object categories in the table of C1–C37 (in order) is: Boeing737, Boeing747, Boeing777, Boeing787, C919, A220, A321, A330, A350, ARJ21, OtherAirplane, SmallCar, Bus, CargoTruck, DumpTruck, Van, Trailer, Tractor, Excavator, TruckTractor, OtherVehicle, BasketballCourt, TennisCourt, FootballField, BaseballField, Intersection, Roundabout, Bridge, PassengerShip, Motorboat, FishingBoat, Tugboat, EngineeringShip, LiquidCargoShip, DryCargoShip, Warship, OtherShip.

Figure 3: Per-category t-SNE visualizations with orientation coloring on FAIR-1M-2.0.

Table 17

Explore similar work

Real-Time Oriented Object Detection Transformer in Remote Sensing Images
Mar 16, 2026Zeyu Ding, Yong Zhou, Jiaqi Zhao +4Object Detection Detection Transformer
School of Computer Science and Technology / School of Artificial Intelligence, China University of Mining and Technology, Xuzhou 221116, China · Mine Digitization Engineering Research Center of the Ministry of Education, Xuzhou 221116, China · Jiangsu Provincial Industrial Technology Engineering Center for Intelligent Sensing and Emergency IoT in Underground Space, Xuzhou 221116, China +1
OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images
Sep 29, 2024Zeyu Ding, Yong Zhou, Jiaqi Zhao +4Remote Sensing Image Understanding Detection Transformer
School of Computer Science and Technology, China University of Mining and Technology, and with Mine Digitization Engineering Research Center of the Ministry of Education, and also with Innovation Research Center of Disaster Intelligent Prevention and Emergency Rescue, China University of Mining and Technology, Xuzhou 221116, China · School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, ON K1N 6N5, Canada
Le-DETR: Revisiting Real-Time Detection Transformer with Efficient Encoder Design
Feb 24, 2026Jiannan Huang, Aditya Kane, Fengzhe Zhou +2Real-Time Object Detection Efficient Object Detection
Beijing Jiaotong University

Mar 16, 2026J/K move · Enter open · S save

Real-Time Oriented Object Detection Transformer in Remote Sensing Images

Zeyu Ding, Yong Zhou, Jiaqi Zhao, Wen-Liang Du+3

School of Computer Science and Technology / School of Artificial Intelligence, China University of Mining and Technology, Xuzhou 221116, China · Mine Digitization Engineering Research Center of the Ministry of Education, Xuzhou 221116, China · Jiangsu Provincial Industrial Technology Engineering Center for Intelligent Sensing and Emergency IoT in Underground Space, Xuzhou 221116, China+1

Object Detection Detection Transformer Real-Time Object Detection Aerial Object Detection

Open paper AskPDF arXiv

Recent real-time detection transformers have gained popularity due to their simplicity and efficiency. However, these detectors do not explicitly model object rotation, especially in remote sensing imagery where objects appear at arbitrary angles, leading to challenges in angle representation, matching cost, and training stability. In this paper, we propose a real-time oriented object detection transformer, the first real-time end-to-end oriented object detector to the best of our knowledge, that addresses the above issues. Specifically, angle distribution refinement is proposed to reformulate angle regression as an iterative refinement of probability distributions, thereby capturing the uncertainty of object rotation and providing a more fine-grained angle representation. Then, we incorporate a Chamfer distance cost into bipartite matching, measuring box distance via vertex sets, enabling more accurate geometric alignment and eliminating ambiguous matches. Moreover, we propose oriented contrastive denoising to stabilize training and analyze four noise modes. We observe that a ground truth can be assigned to different index queries across different decoder layers, and analyze this issue using the proposed instability metric. We design a series of model variants and experiments to validate the proposed method. Notably, our O2-DFINE-L, O2-RTDETR-R50 and O2-DEIM-R50 achieve 77.73%/78.45%/80.15% AP50 on DOTA1.0 and 132/119/119 FPS on the 2080ti GPU. Code is available at https://github.com/wokaikaixinxin/ai4rs.

Abstract

Explore similar work

Real-Time Oriented Object Detection Transformer in Remote Sensing Images

OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images

Le-DETR: Revisiting Real-Time Detection Transformer with Efficient Encoder Design