Beyond Proposals: Region-Consistent Open-Set Object Detection

Jiehui Tang; Yunlong Yu; Zhao Wang; Jungong Han

cs.CVNov 2, 2026

Beyond Proposals: Region-Consistent Open-Set Object Detection

Authors: Jiehui Tang, Yunlong Yu, Zhao Wang, Jungong Han

Organizations: College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, 310027 China · Ningbo Innovation Center, Zhejiang University, Ningbo, China · Department of Automation, Tsinghua University, Beijing 100084

Abstract

Open-set object detection (OSOD) suffers from a fundamental limitation under dense proposal mechanisms: highly overlapping proposals associated with the same object are often assigned contradictory known/unknown predictions. Existing methods perform open-set reasoning independently for each proposal, while ignoring the shared semantics among correlated proposals, resulting in fragmented detections and unstable open-set decisions. In this paper, we reformulate OSOD as an object-centric semantic inference problem, where correlated proposals jointly support a coherent region-level semantic decision. Based on this perspective, we propose OCSI-Net, a unified framework that combines object-centric semantic consistency with geometry-aware uncertainty stabilization. Specifically, OCSI-Net aggregates correlated proposals into structured region clusters and infers region-level semantics from aggregated semantic evidence, thereby enforcing coherent open-set predictions within the same object region. Furthermore, a geometry-aware uncertainty mechanism is introduced to stabilize optimization near open-set transition boundaries. Extensive experiments on multiple benchmarks demonstrate that OCSI-Net consistently achieves state-of-the-art performance, improving the harmonic mean of known and unknown detection by an average of 6.83% over the second-best method.

Figures & tables

Fig. 1 : (a) Proposal prediction: independent decisions over dense RoIs lead to conflicting known (red) and unknown (blue) predictions within the same object region. (b) Structural prediction: aggregating proposals into region clusters resolves semantic conflicts and yields coherent detections.

Fig. 2 : Illustration of the proposed framework during training. The framework first generates dense proposals and subsequently produces proposal-level predictions. The entire detector is optimized under the proposed geometry-aware OpenDet objective, while the Object-Centric Semantic Consistency (OCSC) module is jointly optimized to enforce object-centric semantic coherence among correlated proposals.

Fig. 3 : Illustration of the proposed inference framework. Starting from proposal-level predictions, OCSC groups correlated proposals into object-centric region clusters and performs structured semantic inference based on aggregated region evidence. The inferred region semantics are then used to reconcile inconsistent known/unknown responses, producing coherent and region-consistent open-set detections.

Fig. 4 : Toy illustration of geometry-aware IoU-band reweighting for unknown-probability learning. Colored markers denote proposal features from different known classes, while black triangles indicate unknown proposals. (a) Constant per-proposal weighting treats overlaps equally. (b) Reweighting emphasizes the intermediate-IoU boundary band (inset), yielding more compact local supports and more stable open-set decisions.

	OpenImages-Animal			OpenImages-Vehicle			CUB200			MTSD			Mean $\rm H$
Method	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	Mean $\rm H$
ORE [ 12 ]	37.6	15.6	22.03	33.7	0.3	0.54	53.2	19.8	28.81	41.2	0.4	0.86	13.06
DS [ 18 ]	41.1	15.0	21.89	40.1	2.7	4.71	61.5	21.5	31.79	50.4	5.1	9.23	16.91
VOS [ 6 ]	39.5	16.0	22.74	40.9	9.1	14.59	59.4	8.7	15.16	49.1	4.7	8.46	15.24
Orth.Det [ 27 ]	44.8	8.9	14.13	40.5	5.6	8.17	66.5	24.7	35.92	49.3	3.3	6.07	16.07
FCOS [ 28 ]	30.3	41.8	34.82	30.7	18.7	21.31	53.5	24.7	33.74	41.7	4.4	7.96	24.46

TABLE I : Main comparison results on four benchmarks. We report split-averaged

\mathrm{AP}_{\mathrm{known}}

\mathrm{AP}_{\mathrm{unk}}

, and the harmonic mean

\rm H

(%) for comparison. Results of competitors are from the published literature. The best results are marked in bold .

OpenImages-Animal
	Split1			Split2			Split3			Split4			Mean
Method	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$
ORE [ 12 ]	40.4	17.4	24.32	34.8	13.0	18.93	40.4	19.1	25.94	34.8	13.0	18.93	37.6	15.6	22.03
DS [ 18 ]	44.0	19.0	26.54	36.8	12.3	18.44	43.3	14.0	21.16	40.2	14.6	21.42	41.1	15.0	21.89
VOS [ 6 ]	39.5	17.5	24.25	37.5	13.9	20.28	43.1	14.7	21.92	37.9	18.1	24.50	39.5	16.0	22.74
Orth.Det [ 27 ]	47.0	11.8	18.86	40.5	4.7	8.42	47.2	3.1	5.82	44.5	15.9	23.43	44.8	8.9	14.13

TABLE II : Split-wise OSOD-III [ 11 ] results (%) on OpenImages superclasses. We report

\mathrm{AP}_{\mathrm{known}}

\mathrm{AP}_{\mathrm{unk}}

, and their

\rm H

	Split1			Split2			Split3			Split4			Mean
Method	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$
ORE [ 12 ]	51.3	18.1	26.76	53.6	21.8	30.99	54.4	17.7	26.71	53.6	21.6	30.79	53.2	19.8	28.81
DS [ 18 ]	61.7	19.6	29.75	61.2	22.2	32.58	62.8	22.2	32.80	60.4	21.8	32.04	61.5	21.5	31.79
VOS [ 6 ]	59.7	8.1	14.26	59.5	9.1	15.79	60.5	8.1	14.29	57.7	9.5	16.31	59.4	8.7	15.16
Orth.Det [ 27 ]	64.9	27.2	38.33	67.2	26.9	38.42	67.7	20.9	31.94	66.0	23.8	34.98	66.5	24.7	35.92
FCOS [ 28 ]	55.0	23.0	32.44	55.2	26.1	35.44	50.6	25.0	33.47	53.0	24.6	33.60	53.5	24.7	33.74

TABLE III : Split-wise OSOD-III [ 11 ] results (%) on CUB200. We report

\mathrm{AP}_{\mathrm{known}}

\mathrm{AP}_{\mathrm{unk}}

, and their

\rm H

on four official CUB200 splits.

Method	K ( $\mathrm{AP}_{\mathrm{known}}$ )	U1 ( $\mathrm{AP}_{\mathrm{unk}}$ )	$\rm H$	U2 ( $\mathrm{AP}_{\mathrm{unk}}$ )	$\rm H$	U1+2 ( $\mathrm{AP}_{\mathrm{unk}}$ )	$\rm H$	Mean ( $\mathrm{AP}_{\mathrm{unk}}$ )	Mean ( $\rm H$ )
ORE [ 12 ]	41.2	0.4	0.79	0.2	0.40	0.7	1.38	0.4	0.86
DS [ 18 ]	50.4	4.5	8.26	3.4	6.37	7.5	13.06	5.1	9.23
VOS [ 6 ]	49.1	4.6	8.41	2.9	5.48	6.5	11.48	4.7	8.46
Orth.Det [ 27 ]	49.3	3.9	7.23	1.5	2.91	4.4	8.08	3.3	6.07
FCOS [ 28 ]	41.7	3.8	6.97	3.3	6.12	6.2	10.79	4.4	7.96
Faster R-CNN [ 22 ]	50.0	2.5	4.76	2.3	4.40	4.4	8.09	3.1	5.75

TABLE IV : Split-wise OSOD-III [ 11 ] results (%) on MTSD. Following the benchmark protocol, we report

\mathrm{AP}_{\mathrm{known}}

on the known set (K) and

\mathrm{AP}_{\mathrm{unk}}

on three unknown evaluation sets (U1, U2, and U1+2), and their

\rm H

Settings		Split 1			Split 2			Split 3			Split 4
GABU	OCSC	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$
✗	✗	63.9	23.1	33.93	63.6	30.0	40.80	63.9	26.3	37.26	61.6	28.6	39.06
✓	✗	63.3	30.0	40.71	62.5	39.1	48.11	64.1	34.6	44.94	62.1	37.5	46.76
✗	✓	63.0	29.5	40.18	63.8	37.4	47.16	64.7	34.0	44.58	63.0	34.1	44.25
✓	✓	63.8	35.1	45.29	62.9	45.2	52.60	64.3	41.2	50.22	62.6	40.4	49.11

TABLE V : Ablation results (%) on the CUB200 dataset. We report split-wise

\mathrm{AP}_{\mathrm{known}}

and

\mathrm{AP}_{\mathrm{unk}}

, as well as their harmonic mean

\rm H

Param.	Value	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\rm H$
$\displaystyle u_{\text{high}}$	0.5	63.8	44.7	52.57
	0.7	62.6	42.8	50.84
$\displaystyle u_{\text{mid}}$	0.2	63.1	43.8	51.71
	0.4	63.6	43.8	51.87
$\displaystyle\gamma$	1.0	63.3	44.4	52.19
	1.4	63.2	42.2	50.61

TABLE VI : GABU hyperparameter ablations (%) on CUB200 Split-2. We vary one parameter at a time around the default setting (Base).

\rm H

denotes the harmonic mean of

\mathrm{AP}_{\mathrm{known}}

and

\mathrm{AP}_{\mathrm{unk}}

Panel of Figure 11 — Fig. 5 : Sensitivity of the OCSC semantic decision threshold $\delta$ on CUB200. On two CUB200 splits, we vary $\delta$ by setting it to the $p$ -th percentile of the region-level knownness evidence scores $\{r_{s}\}$ computed on the evaluated split.

Fig. 6 : Qualitative results on four datasets. The competitors produce conflicting known (red) and unknown (blue) predictions within the same region, while OCSI-Net enforces region-consistent semantics and yields coherent detections.

Method / OCSC region encoder	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$	$\mathrm{AP}_{\mathrm{known}}$	$\mathrm{AP}_{\mathrm{unk}}$
	Split 1		Split 2		Split 3		Split 4
DINOv3 [ 24 ] frozen (Ours)	63.8	35.1	62.9	45.2	64.3	41.2	62.6	40.4
DINOv3 [ 24 ] (unfrozen)	64.0	36.5	63.4	45.6	64.5	41.2	62.5	41.0
CLIP-ViT-B/32 [ 21 ] (frozen)	63.7	31.7	62.6	40.5	64.1	37.1	62.2	39.3
CLIP-ViT-B/32 [ 21 ] (unfrozen)	63.7	32.6	62.6	41.8	64.2	37.8	62.3	40.0
DeiT-Tiny [ 29 ] (frozen)	63.7	32.1	62.5	41.5	64.1	35.7	62.1	38.5

TABLE VII : Analysis of different OCSC region encoders on CUB200. We report

\mathrm{AP}_{\mathrm{known}}

and

\mathrm{AP}_{\mathrm{unk}}

on each official split. The default frozen DINOv3 setting is highlighted.

Fig. 7 : Representative failure cases. These examples indicate that region-consistent refinement may be less effective when the region-level evidence is weak or unstable, or when crowded and heavily overlapping scenes require finer-grained instance-level discrimination.

Object Detection Uncertainty Estimation

See all 8 papers

Exploring Hierarchical Consistency and Unbiased Objectness for Open-Vocabulary Object Detection
Apr 25, 2026Sanghoon Lee, Geon Lee, Hyekang Park +1Open-Vocabulary Object Detection Probability Calibration
Yonsei University · Korea Institute of Science and Technology (KIST)
Beyond Known Objects: A Novel Framework for Open-Set Object Detection using Negative-Aware Norm
May 4, 2026Yuchen Zhang, Yao Lu, Johannes BetzAutonomous Driving Perception Open-Set Recognition
TUM School of Engineering and Design, Technical University of Munich, 85748 Garching, Germany; Munich Institute of Robotics and Machine Intelligence (MIRMI)
DeCo-DETR: Decoupled Cognition DETR for efficient Open-Vocabulary Object Detection
Apr 3, 2026Siheng Wang, Yanshu Li, Bohan Hu +12Open-Vocabulary Object Detection Detection Transformer
Jiangsu University · MBZUAI · Brown University +5

Apr 25, 2026J/K move · Enter open · S save

Exploring Hierarchical Consistency and Unbiased Objectness for Open-Vocabulary Object Detection

Sanghoon Lee, Geon Lee, Hyekang Park, Bumsub Ham

Yonsei University · Korea Institute of Science and Technology (KIST)

Open-Vocabulary Object Detection Probability Calibration VLM Adaptation Pseudo-Labeling Object Detection

Open paper AskPDF arXiv

Conventional object detectors typically operate under a closed-set assumption, limiting recognition to a predefined set of base classes seen during training. Open-vocabulary object detection (OVD) addresses this limitation by leveraging vision-language models (VLMs) to generate pseudo labels for novel object classes. However, existing OVD methods suffer from two critical drawbacks: (1) inaccurate class label assignments, as VLMs are optimized for image-level predictions rather than the region-level predictions required for pseudo labeling, and (2) unreliable objectness scores from region proposal networks (RPNs) trained exclusively on base object classes. To address these issues, we propose a novel pseudo labeling framework for OVD. Our approach introduces a hierarchical confidence calibration (HCC) technique, which ensures reliable class label estimation by assessing consistency across hierarchical semantic levels (class, super- and sub-category). We also present LoCLIP, a parameter-efficient adaptation of CLIP that incorporates an objectness token to mitigate base class bias problem of RPNs and provide reliable objectness estimations for novel object classes. Extensive experiments on standard OVD benchmarks, including COCO and LVIS, demonstrate that our approach clearly sets a new state of the art, validating the effectiveness of our approach. Project site: https://cvlab.yonsei.ac.kr/projects/HCC

Abstract

Object Detection Uncertainty Estimation

Explore similar work

Exploring Hierarchical Consistency and Unbiased Objectness for Open-Vocabulary Object Detection

Beyond Known Objects: A Novel Framework for Open-Set Object Detection using Negative-Aware Norm

DeCo-DETR: Decoupled Cognition DETR for efficient Open-Vocabulary Object Detection