Research map / Detection and segmentation / Open-set and few-shot detection
Open-World Objectness Modeling Unifies Novel Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OWOBJ, a class-agnostic probabilistic objectness plugin for DETR/Faster R-CNN detectors that uses a dynamic Gaussian prior in the KL term and an energy-based margin loss to improve unknown-object recall in open-world, few-shot and open-vocabulary detection.
Problem
Object detectors trained on limited labeled data are biased toward known classes, misclassifying unknown objects as known categories or background, which hurts open-world, few-shot and zero-shot detection.
Approach
Model the joint distribution of objectness and class labels via variational approximation, replace the static N(0,1) prior with a dynamic Gaussian N(0, σ²+β²) to avoid log-variance collapse in the low-data regime, and add an energy-based margin loss that pushes unknown objects toward high-density regions, using soft objectness scores as pseudo-labels for the (Z+1)-th classifier dimension.
Key results
- On M-OWODB Task 1 with S-OWODB comparisons, PROB+OWOBJ improves unknown U-Recall from 19.4 to 23.6 (M-OWODB Task 1) and gains about 3% known mAP over PROB.
- On COCO few-shot detection, DeFRCN+OWOBJ improves over DeFRCN by 2.6% at 1-shot and 4.2% at 30-shot novel-class mAP.
- On OV-LVIS, CORA+OWOBJ reaches 31.7 APr versus 28.1 for CORA (+3.6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add OWOBJ to PROB | PROB · M-OWODB Task 1 | U-Recall: 19.4 -> 23.6 | Helped |
| Add OWOBJ to PROB | PROB · M-OWODB Task 1 | known mAP@0.5, current known: 59.5 -> 61.4 | Helped |
| Add OWOBJ to DeFRCN | DeFRCN · COCO few-shot novel classes, 1-shot | mAP: 9.3 -> 11.9 | Helped |
| Add OWOBJ to DeFRCN | DeFRCN · COCO few-shot novel classes, 30-shot | mAP: 22.6 -> 26.4 | Helped |
| Add OWOBJ to CORA | CORA · OV-LVIS rare categories | APr: 28.1 -> 31.7 | Helped |
| Use objectness scores rather than random soft labels for unknown pseudo-labels (ablation) | OWOBJ · M-OWODB Task 1 | U-Recall: 10.2 -> 23.6 | Helped |
| Use objectness scores rather than random soft labels for unknown pseudo-labels (ablation) | OWOBJ · M-OWODB Task 1 | known mAP, current known: 56.1 -> 61.4 | Helped |
| Add objectness loss (ablation) | OWOBJ · M-OWODB Task 1 | U-Recall: 13.4 -> 23.6 | Helped |
| Add KL divergence loss (ablation) | OWOBJ · M-OWODB Task 1 | U-Recall: 19.2 -> 23.6 | Helped |
| Add energy-based margin loss (ablation) | OWOBJ · M-OWODB Task 1 | U-Recall: 21.1 -> 23.6 | Helped |
| Add energy-based margin loss (ablation) | OWOBJ · M-OWODB Task 1 | known mAP, current known: 60.1 -> 61.4 | Helped |
Surprising findings
- Replacing the energy-based and objectness pseudo-labels with random uniform soft labels drops unknown recall by about 50% and causes over 8K misclassifications, showing the objectness model itself, not just soft labels, is needed.
Limitations
- Theoretical analysis of posterior collapse is tied to the Gaussian variational setup and a fixed beta hyperparameter (beta=0.5 best in ablation)
- Gains on S-OWODB and OWOD depend on the PROB/MEPU-FS base pipelines
- Hyperparameters such as energy margin delta=0.2 chosen as defaults
Experimental setup
Datasets: COCO PASCAL VOC 2007 PASCAL VOC 2012 OV-LVIS
Benchmarks: M-OWODB S-OWODB COCO few-shot detection (1-30 shot) OV-LVIS
Metrics: U-Recall known mAP@0.5 COCO-style mAP APr A-OSE
Baselines: ORE OW-DETR PROB MEPU-FS CAT DeFRCN TFA MPSR CORA Detic
Builds on: PROB Deformable DETR CORA DeFRCN Energy-based OOD detection
Closest papers by meaning
- Exploring Orthogonality in Open World Object Detection (CVPR 2024 · similarity 0.92)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.91)
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.91)
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.90)
- Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.89)
- Looking Beyond the Known: Towards a Data Discovery Guided Open-World Object Detection (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.