Class Geometry as Supervision for Sample-Efficient Open-World Detection

Abstract: Open-world object detection requires models to recognize known categories, reject unfamiliar objects, and incorporate new classes over time. This is especially challenging in scarce-data settings such as biomedical and scientific imaging, where rare categories may have only a few annotated examples and fine-grained classes differ by subtle morphology. Prototype-based detectors are natural for this regime, but they typically learn class prototypes as independent anchors, ignoring relational structure among classes. We propose class-geometry supervision (CGS), a general framework that constrains learned prototype or class-representation spaces to preserve visual or semantic class dissimilarities estimated from training data. CGS introduces a dissimilarity-preserving objective that aligns pairwise distances among learned class representations with a target class-geometry matrix while retaining the standard task loss. We instantiate the same objective across prototype recognition, few-shot biomedical object detection, open-set detection, novel-class insertion, and OWOD adaptation on COCO. Experiments show that CGS improves sample efficiency in recognition and ova detection, substantially strengthens novel-class insertion, and improves unknown recall on COCO while retaining much of the known-class detection performance. Ablations show that meaningful visual geometry provides the most reliable gains, while random geometry can help novel separation but is less consistent for few-shot detection. These results suggest that relational class geometry is an effective supervisory signal for building calibrated and extensible open-world detectors under limited supervision.

• The paper introduces a graph-alignment loss that preserves visual or semantic pairwise class dissimilarities in prototype space, improving detector calibration without requiring additional model capacity.
• The method improves parasitic ova detection in few-shot settings, raising 5-shot mAP from 0.066 to 0.076 and novel-only mAP from 0.0629 to 0.1558 during novel-class insertion.
• The approach increases unknown-object recall in COCO open-world detection, but results reveal a tradeoff between stronger unknown rejection and retaining known-class mAP, especially with frozen detector heads.

Class-geometry supervision (CGS) is a relational training objective that constrains the space of learned class prototypes to preserve a target matrix of pairwise class dissimilarities, estimated from visual or semantic information available at training time. The paper develops this idea as a mechanism for sample-efficient open-world detection: detecting known categories from scarce support examples, rejecting objects outside the current label set, and inserting newly labeled classes without large-scale retraining. The framework is instantiated across prototype recognition, few-shot parasitic ova detection, open-set detection, novel-class insertion, and the standard OWOD protocol on COCO, using FSP-DETR and OW-DETR as base detectors.

Motivation and problem setting

The motivating observation is that conventional prototype learning treats class prototypes as independent anchors: each prototype is pulled toward its support examples only through task losses such as matching or classification. Under scarce supervision this leaves the prototype space underconstrained—many prototype configurations explain the limited support data but induce very different behavior for unknown rejection and novel-class insertion. The authors argue that an additional supervisory signal is available even when labels are scarce: classes have geometry, in the form of visual, morphological, or semantic dissimilarities between categories.

Formally, given known classes $\mathcal{C}_t$ with small per-class support sets, prototypes are computed as centroids of region features $z = f_\theta(I,b)$ produced by a frozen-or-trained encoder. A target dissimilarity matrix $D$ is estimated either visually, via average pixel-level MSE between cropped and resized support regions of different classes, or textually, as $1 - \cos(e_i, e_j)$ over MPNet embeddings of Gemini-generated class descriptions. The learned prototype graph $G$, with entries $G_{ij} = d(p_i, p_j)$, is compared against $D$ after off-diagonal standardization of both matrices.

The dissimilarity-preserving objective

The core loss aligns the two complete weighted graphs over the class set:

$$\mathcal{L}_{\mathrm{CG}} = \frac{1}{|\mathcal{E}|} \sum_{(i,j)\in\mathcal{E}} \left(\hat{G}_{ij} - \hat{D}_{ij}\right)^2,$$

added to the task loss with weight $\lambda$: $$\mathcal{L} = \mathcal{L}_{\mathrm{task}} + \lambda \mathcal{L}_{\mathrm{CG}}$$. The objective requires only pairwise relational targets; $z = f_\theta(I,b)$0 need not be a metric. Two propositions characterize what the loss controls. First, by Cauchy–Schwarz, if $z = f_\theta(I,b)$1 then the average absolute distortion between the learned prototype graph and the target geometry is bounded by $z = f_\theta(I,b)$2, so the objective is a genuine graph-alignment criterion rather than an unconstrained regularizer. Second, a margin condition: if the target geometry separates classes $z = f_\theta(I,b)$3 and $z = f_\theta(I,b)$4 relative to class $z = f_\theta(I,b)$5 by margin $z = f_\theta(I,b)$6, and pairwise distortions are each bounded by $z = f_\theta(I,b)$7, then neighborhood order is preserved in the learned space ($z = f_\theta(I,b)$8). This second result underpins the incremental setting: new classes can be inserted into an existing prototype manifold according to their relationships with known classes rather than as isolated anchors.

Open-world decisions use nearest-prototype distance as the unknown score, with a validation-selected threshold $z = f_\theta(I,b)$9. Because CGS calibrates inter-prototype distances, distances from query features to the known manifold become more meaningful for rejection. Notably, the constraint acts on relationships among prototypes rather than on individual embeddings, so intra-class variation is preserved.

Results on ova detection and novel-class insertion

On the parasitic ova benchmark (15 known / 5 held-out classes), generic detectors (DETR, YOLO, Stable-DINO, FCOS) perform near zero in few-shot regimes, while prototype-based methods are stronger. Adding CGS to FSP-DETR improves 5-shot mAP from 0.066 to 0.076 and 10-shot mAP from 0.094 to 0.102, with the largest relative gain in the most data-scarce regime. In open-set detection, mAP improves from 0.045 to 0.061, with a slight mAR decrease attributed to a precision–recall tradeoff: the model becomes more selective in assigning "unknown."

The strongest quantitative results appear in novel-class insertion. Novel-only mAP more than doubles, from 0.0629 to 0.1558, and mAR rises from 0.2595 to 0.3728; generalized-novel mAP improves from 0.0569 to 0.1348. Seen-class retention also improves (mAP 0.0566 → 0.0670), indicating that organizing the prototype space benefits discrimination among previously known categories as well. These gains are obtained without changing the detector backbone, supporting the claim that CGS improves the use of scarce support examples rather than adding capacity.

OWOD adaptation on COCO

On the COCO OWOD protocol, the paper evaluates two regimes against ORE, UC-OWOD, OW-DETR, Fast-OWDETR, OCPL, and RE-OWOD. In the frozen-head setting (fixed OW-DETR features, fine-tuned two-layer MLP head), CGS raises U-Recall above OW-DETR across Tasks 1–3 (e.g., 11.9 vs. 7.5 on Task 1) and even slightly exceeds OW-DETR's Task 1 mAP (62.2 vs. 59.2), but incurs mAP declines on later tasks—for example, Task 2 "Both" drops from 42.9 to 35.6. In the full fine-tuning setting, U-Recall improves on later tasks (9.8 vs. 6.2 on Task 2; 8.2 vs. 5.7 on Task 3) while most known-class mAP figures improve relative to OW-DETR, though Task 2 current-class mAP still declines (28.2 vs. 33.5). Neither variant matches RE-OWOD's overall mAP (33.9 "Both" on Task 4 vs. 32.0 for full-model CGS), so the paper's claim is not state-of-the-art closed-set performance but a controllable mAP–U-Recall tradeoff: full fine-tuning partially recovers known-class discrimination while retaining the open-world benefit. Qualitative examples confirm both directions—recovered known objects and corrected labels, alongside over-prediction of unknown boxes on ambiguous regions and occasional semantic confusion.

Ablations on the source of geometry

All geometry variants improve over the no-CGS baseline, but their reliability differs. Visual morphology $D$0 yields the best few-shot results (0.0759 5-shot, 0.1019 10-shot mAP), outperforming text-derived and combined geometries, indicating that support-crop pixel statistics carry the strongest relational signal for fine-grained ova classification. A notable and somewhat contradictory finding concerns random geometry: a randomly permuted version of $D$1 achieves the highest novel-only mAP (0.1657), suggesting that arbitrary pairwise spreading alone helps separate newly inserted prototypes. However, random geometry gives weaker few-shot gains than visual geometry and lacks interpretable structure, so the paper treats it as evidence that any relational regularization aids separation, while meaningful geometry is required for consistent sample-efficiency gains.

Limitations and open questions

The paper concedes several limitations directly. The known-class mAP cost of frozen-head adaptation is acknowledged as expected but unresolved; full fine-tuning mitigates it without eliminating it (Task 2 current-class mAP remains below baseline). The random-geometry result leaves unclear how much of the novel-insertion gain depends on genuinely meaningful class relationships versus generic spreading of prototypes. Text-derived geometry never surpasses visual geometry in these experiments, and the interaction between the two sources is unexplored beyond simple averaging. Finally, the target geometry $D$2 is assumed fixed and reliable; the authors identify learning or updating class geometry incrementally—and resolving disagreements between visual and semantic geometry—as the main open problem for extending the approach.

Conclusion

This paper reframes prototype-space organization as an explicit supervision problem: pairwise class dissimilarities, estimated cheaply from support crops or text descriptions, are imposed on the learned prototype graph through a standardized squared-distance alignment loss with provable average-distortion and neighborhood-order guarantees. Empirically, the same objective improves low-shot recognition, few-shot and open-set ova detection, produces the largest gains in novel-class insertion (novel-only mAP 0.0629 → 0.1558), and increases unknown recall on COCO OWOD with a tunable tradeoff against known-class mAP. The evidence supports relational class geometry as a practical, architecture-agnostic supervisory signal for calibrated, extensible open-world detectors under limited supervision, contingent on the quality and consistency of the target geometry itself.