PROB: Probabilistic Objectness for Open World Object Detection

1. Introduction

• OWOD requires detecting known objects and novel unknown objects that can be sent for labeling and incrementally learned without forgetting earlier classes.The task is framed as learning over a non-stationary data stream.Source: §1 · p. 2
• Unknown proposals are unlabeled and can be incorrectly penalized as background, leaving existing OWOD methods with ∼10% unknown-object recall.Prior methods use heuristics such as pseudo-labeling to distinguish unknown objects from background.Source: §1 · p. 2
• PROB adds a probabilistic objectness head to deformable DETR, alternating distribution estimation with likelihood maximization for known objects.At inference, the estimated distribution scores whether each proposal is an object; the approach does not require negative examples.Source: §1 · p. 2
• On MS-COCO, PROB reports 100-300% relative gains in unknown recall and ∼10% gains in known object detection across tasks.The paper also reports that PROB outperforms existing OWOD methods on its benchmarks.Source: §1 · p. 2
• In separate incremental-learning experiments, PROB outperformed both OWOD baselines and baseline incremental-learning methods.The paper states that its code and model will be open-sourced upon publication.Source: §1 · p. 2

Open World Object Detection

• ORE combines feature-space contrastive clustering, an RPN-based unknown detector, and an Energy Based Unknown Identifier; later work explored class-distribution separation and localization-based objectness.Yu et al. [33] reduced confusion between known and unknown objects, while Wu et al. reported gains in unknown recall.Source: §2 · p. 2
• OW-DETR uses high-backbone-activation unmatched proposals as pseudo-labeled unknowns, while MViTs were reported to improve ORE unknown detection by ∼4× with additional supervision and larger datasets.The MViT work motivates the potential of transformer-based models, though it does not follow the OWOD objective as described.Source: §2 · p. 2
• Unlike prior methods that estimate objectness, PROB probabilistically estimates it and integrates it directly into class prediction to improve unknown-object detection.The cited prior work includes [9].Source: §2 · p. 3

Class Agnostic Object Detection

• Class-agnostic object detection learns general objectness from limited labeled classes; BackErase and localization losses address sparse labels, but direct integration of Kim et al.’s method into ORE produced a 70% drop in unknown recall.Localization losses avoid penalizing false positives, while PROB indirectly uses this insight.Source: §2 · p. 3

Problem Formulation

• Standard object-detection training uses N images with labels for K known classes, where each object label contains a class vector and bounding-box coordinates.The training dataset is denoted D = {I, Y}.Source: §3 · p. 3
• In the open-world formulation, unknown classes of interest are sent to an oracle for labeling, then added in a later task while a limited subset of earlier data may be retained.The model updates from the new task data, its previous model, and that limited earlier data.Source: §3 · p. 3

DETR for Open World Learning

• DETR maps learned queries to image-specific embeddings and object predictions, but OWOD must distinguish unlabeled unknown objects from background; OW-DETR trains on top unmatched queries selected by backbone activation.OW-DETR selects the top five unmatched queries as unknown pseudo-labels and leaves inference unchanged.Source: §3 · p. 3

4. Method

• PROB adapts deformable DETR (D-DETR) [37] for open-world detection by adding a probabilistic objectness head.The method also uses objectness scores when selecting exemplars for incremental learning.Source: §4 · p. 3, §4.2 · p. 5

4.1. Probabilistic Objectness

• PROB separates objectness from class prediction, multiplying the objectness probability by class probabilities to form final predictions.The objectness head estimates whether a query represents an object, while the classification head predicts its known class.Source: §4.1 · p. 4, Figure 2 · p. 4
• Training alternates between estimating a class-agnostic Gaussian distribution over query embeddings and maximizing the likelihood of matched embeddings.Distribution parameters use exponentially averaged empirical means and covariances; the objectness loss penalizes matched embeddings’ squared Mahalanobis distance.Source: §4.1 · p. 5

4.2. Objectness for Incremental Learning

• For incremental learning, PROB selects the top and bottom 25 objectness-scoring matched objects per class as exemplars.Low-objectness examples are expected to be difficult, while high-objectness examples are expected to represent their classes; exemplar counts are randomly subsampled to match prior work when needed.Source: §4.2 · p. 5

5. Experiments & Results

• Evaluation spans M-OWODB and S-OWODB, using U-Recall for unknown detection and mAP for known classes; the detector uses deformable DETR with a DINO-pretrained ResNet-50 FPN.M-OWODB introduces 20 classes per task, while S-OWODB preserves super-category separation across tasks.Source: §5 · p. 5

5.1. Open World Object Detection Performance

• Across both OWOD benchmarks, PROB reaches 2-3x previous methods’ U-Recall and outperforms existing models on known-object mAP.The comparison covers M-OWODB and S-OWODB; Task 4 has no U-Recall because all classes are known.Source: §5.1 · p. 6, Table 1 · p. 6
• On M-OWODB Tasks 1–3, PROB’s A-OSE was 5195, 6452, and 2641, versus OW-DETR’s 10240, 8441, and 6803.A-OSE counts unknown objects classified as known.Source: §5 · p. 5, §5.1 · p. 7

5.2. Ablation Study

• Without likelihood maximization, PROB-Obj slightly raises unknown recall but sharply reduces known-object mAP; without active exemplar selection, unknown-recall gains are smaller.The exemplar-selection ablation has no effect in Task 1 because no exemplar replay is used there.Source: Table 2 · p. 6, §5.2 · p. 7
• Across Tasks 2–3, PROB detects the skateboard as unknown and later classifies it, while OW-DETR’s example forgets the previously known suitcase.The table’s oracle comparison also suggests that learning unknowns rather than treating them as background can improve known-object detection.Source: Figure 4 · p. 8, §5.2 · p. 8

5.3. Incremental Object Detection

• On PASCAL VOC, PROB’s final mAP was 66.5, 70.1, and 72.6, versus OW-DETR’s 65.7, 69.4, and 70.2, respectively.The three settings introduce 10, 5, or 1 additional classes after initial training on 10, 15, or 19 classes.Source: §5.3 · p. 8, Table 3 · p. 8

6. Conclusions

• PROB integrates probabilistic objectness into deformable DETR and significantly improves unknown detection across OWOD benchmarks.The conclusion also reports that PROB significantly outperforms existing OWOD methods on all benchmarks.Source: §6 · p. 8
• The authors identify substantial room for improvement in unknown detection and other aspects of OWOD.Source: §6 · p. 8
• The conclusion frames probabilistic models as having potential for algorithms that operate reliably in the open world.Source: §6 · p. 8

• WI and A-OSE increase monotonically as objectness temperature rises.Source: Figure 5 · p. 11
• U-Recall and AP50 peak at an objectness temperature of 1.0–1.6.Source: Figure 5 · p. 11
• The sweep evaluates model performance on Task 1 and Task 2.Source: Figure 5 · p. 11

A.1. Evaluation using WI and A-OSE Metrics

• U-Recall measures detection of unknown instances, while WI and A-OSE measure confusion when unknowns are predicted as known classes.WI can rise as U-Recall grows even if unknown-object precision stays the same; A-OSE is less affected by U-Recall.Source: §A.1 · p. 11
• PROB outperforms other OWOD methods in U-Recall while maintaining lower or similar A-OSE.For Task 2, temperature 0.5 reduces A-OSE to 562 and WI by almost 50% to 0.0133, with only marginal known-mAP degradation.Source: §A.1 · p. 11

A.2. Incremental Learning

• Active selection of high- and low-objectness exemplars further improves unknown and known detection performance during incremental learning.Known-object performance gains grow across Tasks 2–4, with deltas of 0.5, 0.7, and 0.9, respectively.Source: §A.2 · p. 11

• Objectness temperature can be changed at test time without affecting training, and higher temperature yields less confident objectness predictions.All experiments used temperature 1.3; U-Recall and known mAP have optimum points that do not necessarily coincide, while WI and A-OSE trade off against U-Recall and mAP.Source: §A.3 · p. 11
• PROB achieves improved WI, A-OSE, and U-Recall over OW-DETR across M-OWODB tasks.Source: Table 4 · p. 12
• PROB achieves favorable performance against existing OWOD approaches in all three PASCAL VOC incremental-detection settings.Source: Table 5 · p. 12
• For methods with lower WI and A-OSE in Table 4, some objectness temperature lets PROB outperform them on known mAP, U-Recall, WI, and A-OSE.Source: §A.3 · p. 12

• PROB detects frisbees as unknown in Tasks 1 and 2, then as a known class in Task 3 after they are added.OW-DETR detects only one frisbee in Task 1, none in Task 2, and both as frisbees in Task 3.Source: B. Additional Qualitative Results · p. 12
• PROB's unknown-object confidence is comparable to its known-object confidence, while OW-DETR's unknown detections tend to have very low confidence.The qualitative examples report OW-DETR unknown detections near 10% confidence.Source: B. Additional Qualitative Results · p. 12, Figure 6 · p. 13
• PROB retains unknown detections across later tasks, while OW-DETR can forget previously detected objects.In one example, OW-DETR forgets the remote, book, and trash can in Task 3 after detecting them in Tasks 1 and 2.Source: B. Additional Qualitative Results · p. 13, Figure 7 · p. 14
• PROB detects unknown objects more consistently across tasks and instances, including multiple objects from the same class.Examples include laptop and keyboard detections across tasks and zebras, giraffes, and kites across instances; OW-DETR misses some instances, including a foreground zebra.Source: B. Additional Qualitative Results · p. 13

C. Additional Implementation Details

• Assuming iid-distributed channels did not change performance, improved model stability, and reduced training time; an unrestricted covariance became diagonal under the regularization needed for stable training.The iid assumption also makes covariance inversion trivial to compute during training.Source: C. Additional Implementation Details · p. 15
• The objectness temperature τ was set to 1.3 in all experiments based on temperature-sweep experiments.The temperature is used for objectness prediction during inference and evaluation.Source: C. Additional Implementation Details · p. 15
• Training used four Nvidia A100 40GB GPUs with batch size 5, a 2 × 10−3 learning rate, and a drop after 35 epochs; incremental-learning finetuning reduced the learning rate tenfold.Other hyperparameters followed OW-DETR [9].Source: C. Additional Implementation Details · p. 15

D. Limitations and Social Impacts

• Current benchmarks expose an entire novel-object dataset per task; the authors propose adding active learning to discover only unknown objects detected by the model.They also note that known and unknown predictions receive different supervision, leading to imbalanced prediction-score and confidence relations; energy-based models are suggested for exploration.Source: D. Limitations and Social Impacts · p. 15
• Unknown-object detection may be biased by the initial training dataset, so future work should assess bias and break detection results down by relevant subclasses.The authors frame these as concerns for OWOD systems bridging benchmarks and real-world settings.Source: D. Limitations and Social Impacts · p. 15
• Real-world exemplar images may contain private or sensitive information about unknown classes, motivating replacement or censoring of stored exemplars.The authors also call for forgetting capabilities that can target particular classes for legal or privacy concerns.Source: D. Limitations and Social Impacts · p. 15