Accurate oriented detection requires a box to follow object boundaries, yet a standard regression head predicts its center, size, and angle from a pooled region representation. Small errors in these coupled parameters can displace long boundaries even when the object is correctly classified. We propose NIF, a refinement module that replaces an additional box regressor with normal-intersection fields. Given a coarse oriented box, the module observes local features at 20 boundary anchors and predicts signed distances along their outward normals. These observations admit an exact affine representation in five geometric variables. A differentiable, reliability-weighted least-squares fit projects the predicted field onto this representation and analytically recovers a refined rectangle. Tangential pooling preserves the normal position profile, while shared projection limits the added network to 0.132M parameters. On DOTA-v1.0, NIF reaches 78.00%, 79.11%, and 79.02% mAP with Oriented R-CNN, LSKNet-S, and Strip R-CNN, respectively, achieving the highest single-scale SOTA result in Table 1. It reaches 68.19% on DIOR-R, improves matched FAIR1M/DIOR-R baselines by 0.30/0.68%, and raises HRSC mean IoU from 0.867 to 0.888 while reducing field error by 24.5%. Across four different detector frameworks with the detectors frozen, relative AP75 improves by 1.57-4.18% and COCO-style mAP by 0.42-3.74%, without detector retraining. Code is available at https://github.com/zsk466991125-collab/Normal-Intersection-Fields.
Figures & tables
Figure 1: Overview of NIF. The carrier predicts coarse boxes P and class scores p . Shared projection reduces FPN features to 64 channels; twenty 3×9 strips predict normal observations and reliabilities. Weighted fitting and analytic decoding refine boxes before rotated NMS. The DOTA example shows raw intersections q^i and compatible intersections qi+ , with signed normal distances in pixels. GT supplies supervision and visualization. Here, JP=J(P) , N=4m=20 , and rˉ is mean observation reliability. Lobsw denotes the weighted loss in Eq. 11 . In the decoder tuple, w+=2a+ and h+=2b+ are full refined box dimensions.
Figure 2: Geometry of the normal-intersection field. (a) Anchors and outward normals are attached to the coarse rectangle. (b) The signed displacement ends at a corresponding target supporting line, including intersections beyond finite edges. (c) Translation, scale, and rotation induce different, coordinated constant and slope patterns. These patterns motivate an exact five-dimensional compatible field. Intersection labels qi denote target points qi∗ ; c,u,v define the coarse-box frame.
Figure 3: Local observation network. Tensor sizes are per observation. T-pool averages along the tangent while preserving nine normal positions; C concatenates visual and geometric descriptors. The shared predictor outputs 20 field values and reliabilities per box. WLS fits five geometric variables for analytic rectangle decoding.
Method
PL
BD
BR
GTF
SV
LV
SH
TC
BC
ST
SBF
RA
HA
SP
HC
mAP
R. Faster R-CNN [ 40 ]
89.40
81.81
47.28
67.44
73.96
73.12
85.03
90.90
85.15
84.90
56.60
64.77
64.70
70.28
62.22
73.17
R 3 Det [ 51 ]
89.50
81.20
50.50
66.10
70.90
78.70
78.20
90.80
85.30
84.20
61.80
63.80
68.20
69.80
67.20
73.70
R 3 Det-GWD [ 53 ]
88.82
82.94
55.63
72.75
78.52
83.10
87.46
90.21
86.36
85.44
64.70
61.41
73.46
76.94
57.38
76.34
R 3 Det-KLD [ 56 ]
88.90
84.17
55.80
69.35
78.72
84.08
87.00
89.75
84.32
85.73
64.74
61.80
76.62
78.49
70.89
77.36
S 2 A-Net [ 16 ]
89.11
82.84
48.37
71.11
78.11
78.39
87.25
90.83
84.90
85.64
60.36
62.60
65.26
69.13
57.94
74.12
RoI Transformer [ 9 ]
89.01
77.48
51.64
72.07
74.43
77.55
87.76
90.81
79.71
85.27
58.36
64.11
76.50
71.99
54.06
74.05
Table 1: Comparisons with SOTA methods on the DOTA-v1.0 dataset (AP in percent). PL, BD, BR, GTF, SV, LV, SH, TC, BC, ST, SBF, RA, HA, SP, and HC follow the standard DOTA category abbreviations.
Method
Params (M)
GFLOPs
AP 50
AP 75
mAP
PASCAL VOC 2007
O-RCNN [ 47 ]
41.37
211.43
90.40
88.60
70.66
O-RCNN + NIF
41.50
220.48
90.50
89.70
74.33
LSKNet-S [ 28 ]
30.98
173.68
90.50
88.20
69.80
LSKNet-S + NIF
31.11
182.72
90.30
89.40
74.55
PKINet-v2-S [ 2 ]
30.73
173.68
90.30
88.40
69.53
Table 2: Comparisons with SOTA methods on the HRSC2016 dataset (AP in %). All models are trained locally on trainval and evaluated on test with the final 36-epoch checkpoint. Complexity uses 1024×1024 reference inputs and the common mean NIF cost described in the text. AP uses VOC 2007/2012 interpolation; mAP follows the COCO-style metric.
Method
G. V. * [ 49 ]
RetinaNet * [ 30 ]
C-RCNN * [ 3 ]
F-RCNN * [ 40 ]
RoI Trans. * [ 9 ]
O-RCNN [ 47 ]
LSKNet-S [ 28 ]
Strip R-CNN [ 62 ]
O-RCNN + NIF
LSKNet-S + NIF
Strip R-CNN + NIF
mAP (%)
29.92
30.67
31.18
32.12
35.29
45.60
47.87
47.92
46.36
48.13
48.22
Table 3: Comparisons with SOTA methods on the FAIR1M-v1.0 dataset . *: Results are referenced from the FAIR1M paper [ 42 ] .
Method
Params (M)
GFLOPs
mAP (%)
RetinaNet-O [ 30 ]
–
–
57.55
Faster R-CNN-O [ 40 ]
41.10
198.00
59.54
TIOE-Det [ 36 ]
41.10
198.00
61.98
O-RCNN [ 47 ]
41.37
211.43
64.30
ARS-DETR [ 63 ]
41.10
198.00
66.12
O-RepPoints [ 26 ]
36.60
–
66.71
Table 4: Comparisons with SOTA methods on the DIOR-R dataset . Literature complexity follows Strip R-CNN [ 62 ] where reported. O-RCNN, LSKNet-S, Strip R-CNN, and their NIF counterparts use the standardized estimates in Table 2 : 1024×1024 input, 1,000 baseline proposals, and the common mean NIF cost.
Frozen carrier
Carrier + NIF
Carrier
AP 75
mAP
AP 75
mAP
YOLO26-L OBB [ 24 ]
59.19
54.18
60.12
54.41
O2-RTDETR [ 13 ]
54.28
51.09
56.05
52.10
LSKNet-S + FAA [ 28 , 15 ]
51.88
48.35
54.05
50.16
PKINet-v2-S [ 2 ]
54.58
50.83
56.17
52.07
Table 5: Cross-framework transfer on DOTA-v1.0 . Each carrier is initialized from its official single-scale checkpoint, with all detector parameters frozen and only NIF optimized. AP 75 is evaluated at IoU 0.75 ; mAP is the COCO-style mean AP averaged over IoU thresholds from 0.50 to 0.95 . “LSKNet-S + FAA” denotes the FAA Head carrier.
Figure 4: Geometric refinement evidence on HRSC2016. (a) Normal observations, box refinement, and Grad-CAM for the final update. (b) Normal-field error. (c) Paired IoU gain by coarse IoU.
Class
Strip R-CNN
O-RCNN+NIF
LSKNet-S+NIF
Strip R-CNN+NIF
Airplane
71.90
71.59
80.84
71.88
Airport
35.66
40.80
44.48
41.55
Baseball field
80.23
72.00
80.24
79.95
Basketball court
88.69
89.29
89.62
89.57
Bridge
44.89
43.43
45.18
45.19
Chimney
78.23
77.41
78.30
78.73
Table 6: DIOR-R per-class AP (%). Native DIORDataset evaluation at rotated IoU 0.5 with VOC2007 eleven-point AP.
Figure 5: Qualitative DIOR-R test-set refinements from LSKNet-S+NIF (epoch 12). White dashed, orange dashed, and green solid outlines represent the matched ground truth, the coarse detector box, and the NIF-refined box. Each panel reports the rotated IoU before and after refinement for the same candidate.
Coarse
Fine class
GV ∗
Retina ∗
Cascade ∗
Faster ∗
ROI Trans. ∗
O-RCNN
LSKNet-T
LSKNet-S
Strip R-CNN
O-RCNN+NIF
LSKNet-S+NIF
Strip R-CNN+NIF
Airplane
Boeing737
35.43
38.46
40.42
36.43
39.58
42.84
45.12
39.84
40.08
44.58
40.96
43.75
Boeing747
47.88
55.36
52.86
50.68
73.56
87.61
84.97
86.63
85.82
85.94
85.58
86.29
Boeing777
15.67
24.75
29.07
22.50
18.32
18.83
20.16
24.21
18.73
19.57
22.18
22.80
Boeing787
48.32
51.81
52.47
51.86
56.43
62.92
56.00
56.48
54.97
60.36
54.70
61.51
C919
0.01
0.81
0.00
0.01
0.00
22.17
25.77
24.17
23.12
20.99
28.07
27.32
A220
40.11
40.50
44.37
47.81
47.67
47.87
50.05
52.20
49.04
48.59
51.16
50.62
Table 7: FAIR1M-v1.0 fine-grained AP (%).
Figure 6: Qualitative FAIR1M-v1.0 test-set predictions from LSKNet-S+NIF (epoch 12).
RiO-DETR is the first real-time end-to-end transformer for oriented object detection, yet further localization gains require a more expressive box representation. Although distributional regression has succeeded in horizontal DETRs, extending it to oriented detection introduces geometry-misaligned boundary modeling and topology-inconsistent orientation refinement. We present RiO-DETRv2, a geometry-aware uncertainty modeling framework for real-time oriented object detection. To resolve the geometry mismatch, Object-Aligned Boundary Distribution Refinement represents box boundaries as probability distributions in an object-aligned coordinate system and progressively refines them across decoder layers. For better orientation topology consistency, we introduce Circular-to-Local Angular Residual Refinement that transitions from global periodic search to local fine-grained estimation. Unified OBB Quality-Aligned Learning further aligns classification, boundary and angular refinement with complete-OBB localization quality. Together, these designs improve localization without compromising efficiency, achieving a better balance between speed and accuracy. Specifically, RiO-DETRv2-l/x achieves 53.7/54.1 mAP50-95 on DOTA, outperforming RiO-DETR-l/x by 1.2/1.5 points with nearly unchanged computational cost. Code is available at https://github.com/RicePasteM/RiO-DETRv2.
Zhangchi Hu, Zilong He, Yifan Zhao +6
University of Science and Technology of China · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center
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.
Zhangchi Hu, Yifan Zhao, Yansong Peng +8
University of Science and Technology of China · Tsinghua University · Huawei Technologies Co., Ltd. +1
With the increasing availability of high-resolution remote sensing and aerial imagery, oriented object detection has become a key capability for geographic information updating, maritime surveillance, and disaster response. However, it remains challenging due to cluttered backgrounds, severe scale variation, and large orientation changes. Existing approaches largely improve performance through multi-scale feature fusion with feature pyramid networks or contextual modeling with attention, but they often lack explicit foreground modeling and do not leverage geometric orientation priors, which limits feature discriminability. To overcome these limitations, we propose FGAA-FPN, a Foreground-Guided Angle-Aware Feature Pyramid Network for oriented object detection. FGAA-FPN is built on a hierarchical functional decomposition that accounts for the distinct spatial resolution and semantic abstraction across pyramid levels, thereby strengthening multi-scale representations. Concretely, a Foreground-Guided Feature Modulation module learns foreground saliency under weak supervision to enhance object regions and suppress background interference in low-level features. In parallel, an Angle-Aware Multi-Head Attention module encodes relative orientation relationships to guide global interactions among high-level semantic features. Extensive experiments on DOTA v1.0 and DOTA v1.5 demonstrate that FGAA-FPN achieves state-of-the-art results, reaching 75.5% and 68.3% mAP, respectively.
Jialin Ma
International school of BUPT, Beijing University Of Posts And Telecommunications, Beijing,China.