Detecting Unknown Objects via Energy-Based Separation for Open World Object Detection

Conference: CVPR 2026 arXiv: 2603.29954 Code: N/A Area: Object Detection Keywords: Open World Object Detection, Energy Function, Unknown Object Detection, Incremental Learning, Catastrophic Forgetting

TL;DR

This paper proposes the DEUS framework, which introduces ETF-Subspace Unknown Separation (EUS) to effectively separate known, unknown, and background proposals via energy scores within geometrically orthogonal known/unknown subspaces, and designs an Energy-based Known Distinction (EKD) loss to reduce cross-task interference between old and new classes during incremental learning, achieving substantial improvements in unknown object recall on OWOD benchmarks.

Background & Motivation

Open World Object Detection (OWOD) is a highly challenging setting that requires detectors to: 1. Incrementally learn known classes: continuously expand the set of recognizable categories; 2. Detect unknown objects: identify unseen objects without annotations; 3. Avoid catastrophic forgetting: retain knowledge of old classes when learning new ones.

Existing OWOD methods suffer from two core problems:

Problem 1: Insufficient representation learning for unknown objects - Existing methods (including energy-based ones) rely heavily on the detector's known-class predictions to detect unknown objects. - Energy is modeled only in the known space, pushing non-known objects away from known regions, but without constraints to prevent unknown objects from being confused with background. - Result: Known, unknown, and background features are entangled in feature space, causing many unknowns to be missed or misclassified.

Problem 2: Cross-task interference between old and new classes during memory replay - Memory replay alleviates forgetting of old classes, but lacks explicit regularization to prevent mutual interference between old and new classes. - Interference worsens as more tasks and categories are added. - Result: A trade-off exists between retaining old-class knowledge and learning new classes.

The two designs in DEUS correspond to addressing these two problems respectively.

Method

Overall Architecture

Built upon OrthogonalDet as the base detector, with two additional modules: - EUS (ETF-Subspace Unknown Separation): constructs orthogonal known/unknown subspaces to guide the separation of proposal features; - EKD (Energy-based Known Distinction): separates energy responses of old and new classifiers during memory replay.

Key Designs

1. ETF-Subspace Unknown Separation (EUS):

   • Constructing orthogonal subspaces: Simplex ETF (Equiangular Tight Frame) is used to generate \(K\) equiangular basis vectors, divided into a known subspace \(W_\mathcal{K}^E\) (first \(K/2\) vectors) and an unknown subspace \(W_\mathcal{U}^E\) (last \(K/2\) vectors), with \(K=128\).
   • ETF basis vectors are fixed and non-learnable, guaranteeing geometric orthogonality between the two subspaces.
   • Dual-space energy computation: For each proposal feature \(f\), the Helmholtz free energy is computed separately in the two subspaces: $\(E^\mathcal{K}(f) = -\log \sum_{i=1}^{K/2} \exp(W_{\mathcal{K},i}^E \cdot f)\)$ $\(E^\mathcal{U}(f) = -\log \sum_{i=1}^{K/2} \exp(W_{\mathcal{U},i}^E \cdot f)\)$
   • Unknown offset: \(\Delta_u(f) = s_u(f) - s_k(f)\) (unknown score minus known score); a positive value indicates a higher likelihood of being unknown.
   • Learning objectives: known proposals → \(\Delta_u \leq -m\); unknown proposals → \(\Delta_u \geq m\); background → boundary region.
   • Dual loss: energy margin loss on \(\Delta_u\) + subspace focal loss; the former provides the primary separation mechanism, while the latter stabilizes training.
   • Inference-time calibration: subspace information is injected into the detector's existing unknown logit: \(z_u' = z_u + \sigma_{z_u} \tilde{\Delta}_u(f)\), where \(\tilde{\Delta}_u\) is the normalized offset.

2. Energy-based Known Distinction (EKD):

   • Split classifier: the known-class classification head is split into an old-task sub-classifier \(H_{prev}\) and a new-task sub-classifier \(H_{curr}\).
   • Energy score: \(S(f;H) = \log \sum_{c=1}^{C_H} \exp(z_c(f;H))\); higher values indicate stronger affinity with the corresponding classifier.
   • Contrastive loss: encourages old-class proposals to score high on the old classifier and low on the new classifier, and vice versa: $\(\mathcal{L}_{prev} = \log(1 + \exp[S(f_{prev};H_{curr}) - S(f_{prev};H_{prev})])\)$ $\(\mathcal{L}_{curr} = \log(1 + \exp[S(f_{curr};H_{prev}) - S(f_{curr};H_{curr})])\)$
   • Activated only during memory replay (i.e., incremental task training).

Loss & Training

Total loss: $\(\mathcal{L}_{total} = \mathcal{L}_{cls} + \mathcal{L}_{bbox} + \mathcal{L}_{EUS} + \mathcal{L}_{EKD}\)$

• \(\mathcal{L}_{cls}\): sigmoid focal loss
• \(\mathcal{L}_{bbox}\): L1 + GIoU loss
• \(\mathcal{L}_{EUS} = \mathcal{L}_{energy} + \mathcal{L}_{subspace}\) (weight 1.0)
• \(\mathcal{L}_{EKD}\) (weight 1.0, enabled only during incremental tasks)
• ETF space dimension \(K=128\) (64 vectors each for known and unknown subspaces)
• Implemented based on MMDetection
• Improved pseudo-label strategy: dynamically scales pseudo-label count and filters noisy detections

Key Experimental Results

Main Results

M-OWODB Benchmark:

Method	T1 U-Rec	T1 H-Score	T2 U-Rec	T2 H-Score	T3 U-Rec	T3 H-Score	T4 Known mAP
OrthogonalDet	36.3	46.6	30.2	38.0	28.7	35.7	44.7
O1O	49.3	56.1	50.3	51.6	49.5	47.4	42.4
DEUS	65.1	65.6	66.2	59.0	69.0	58.0	46.0

S-OWODB Benchmark:

Method	T1 U-Rec	T1 H-Score	T2 U-Rec	T3 U-Rec	T4 Known mAP
OrthogonalDet	24.6	36.6	27.9	31.9	46.2
O1O	49.8	59.1	51.1	48.1	45.9
DEUS	68.7	70.1	62.9	60.7	48.8

DEUS nearly doubles U-Rec (e.g., M-OWODB T1: 36.3→65.1) while maintaining competitive known mAP.

RS-OWODB (Remote Sensing):

Method	T1 H-Score	T2 H-Score	T3 H-Score	T4 mAP
OrthogonalDet	34.8	15.6	16.2	64.2
DEUS	62.5	39.4	40.9	68.3

Ablation Study

EUS	EKD	T1 U-Rec	T1 H-Score	T2 Known mAP	T3 H-Score	T4 Known mAP	Note
✗	✗	36.8	47.2	52.0	37.6	44.7	Baseline
✗	✓	36.8	47.2	52.6	43.9	45.9	EKD improves known
✓	✗	65.1	65.6	51.9	57.5	43.5	EUS substantially improves U-Rec
✓	✓	65.1	65.6	53.3	58.0	46.0	Best combined

Key Findings

1. EUS is critical for unknown object detection: U-Rec jumps from 36.8 to 65.1, nearly doubling, demonstrating that dual-subspace modeling is far superior to using only the known space.
2. EKD independently improves known-class performance: regardless of whether EUS is present, EKD consistently improves mAP across tasks.
3. The two modules are complementary: EUS boosts unknown detection while EKD protects known performance; their combination achieves the best H-Score across all settings.
4. Negligible computational overhead: inference time increases by only 1.9%, FLOPs by +0.5%, and training time by +6.2%.
5. Generalization to remote sensing: H-Score on RS-OWODB improves from 34.8 to 62.5, demonstrating that the method is not limited to natural images.
6. PCA visualization clearly shows that known/unknown/background features are severely entangled in the baseline, while DEUS achieves clean three-way separation.

Highlights & Insights

1. Dual-subspace energy modeling: DEUS is the first work in OWOD to explicitly model a representation space for unknown objects, rather than merely excluding them from the known space. The orthogonality guaranteed by ETF is key—it prevents the two spaces from overlapping.
2. Unified use of energy functions: energy functions serve dual purposes—known/unknown separation (EUS) and old/new class distinction (EKD)—forming a unified energy-based framework.
3. Geometric advantages of ETF: the equiangular tight frame provides fixed, uniformly distributed orthogonal basis vectors, ensuring spatial separation without learning.
4. Calibrated inference: injecting the subspace offset into existing logits is a simple and efficient approach that does not disrupt the detection pipeline.
5. Improved pseudo-labels: dynamic scaling and noise filtering serve as auxiliary contributions that also provide practical improvements to baseline performance.

Limitations & Future Work

1. Semantic overlap between known and unknown: the paper acknowledges that separation remains difficult when known and unknown categories are semantically similar.
2. ETF dimension selection: \(K=128\) is a hyperparameter that may require tuning for different datasets.
3. EUS may slightly reduce known mAP: ablation results show that using EUS alone causes T4 Known mAP to drop from 44.7 to 43.5, as more proposals are labeled as unknown.
4. Validated only within the Faster R-CNN paradigm: applicability to DETR-based open-world detectors remains unknown.
5. Pseudo-label quality has room for improvement: the current approach relies on a dynamic matcher; self-training or consistency regularization could be considered.

Related Work & Insights

• OrthogonalDet: the base model, which decouples objectness and classification predictions via orthogonalization.
• PROB (Zohar et al.): models class-agnostic objectness using a Gaussian distribution.
• Du et al. (Unknown-Aware OD): energy uncertainty regularization, but only in the known space.
• Neural Collapse / ETF: the geometric properties of ETF in classification are cleverly leveraged to construct separated subspaces.
• Inspiration: the dual-subspace + energy paradigm may generalize to other open-world tasks (e.g., open-world segmentation, open-world tracking).

Rating

• Novelty: ⭐⭐⭐⭐⭐ — ETF dual-subspace + energy separation is a novel design; EKD's energy-based distinction between old and new classes is also an innovative contribution.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Three benchmarks (M-OWODB/S-OWODB/RS-OWODB), comprehensive ablations, analysis, and visualizations.
• Writing Quality: ⭐⭐⭐⭐ — Motivation and methodology are clearly articulated, though the notation is dense and requires careful reading.
• Value: ⭐⭐⭐⭐⭐ — Nearly doubling U-Rec for unknown detection in OWOD is a significant contribution with practical impact on open-world learning.

Related Papers

• [CVPR 2026] EW-DETR: Evolving World Object Detection via Incremental Low-Rank DEtection TRansformer
• [CVPR 2026] Show, Don't Tell: Detecting Novel Objects by Watching Human Videos
• [CVPR 2026] Beyond Prompt Degradation: Prototype-Guided Dual-Pool Prompting for Incremental Object Detection
• [CVPR 2026] Parameter-Efficient Semantic Augmentation for Enhancing Open-Vocabulary Object Detection
• [CVPR 2026] NoOVD: Novel Category Discovery and Embedding for Open-Vocabulary Object Detection