Research map / Detection and segmentation / Open-set and few-shot detection
Exploring Orthogonality in Open World 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 orthogonality-based open world object detection on RandBox: polar feature decomposition (magnitude for objectness, direction for class), a squared-correlation decorrelation loss, and a routed cross-task calibration layer for incremental learning.
Problem
Open world object detection suffers from low recall on unknown objects and misclassification of unknown objects into known classes, because objectness predictions are correlated with class information; incremental learning also causes interference with known classes.
Approach
Enforce orthogonality at three levels: decouple objectness (feature magnitude) and class (feature direction) via a polar coordinate decomposition, add a squared Pearson-correlation loss between objectness and class predictions, and use a Gumbel-routed task-specific affine calibration layer to keep orthogonality under incremental task changes.
Key results
- On M-OWODB Task 2, U-Recall of 26.3 vs 6.4 for the base model and up to 7.3% above prior best CAT.
- On M-OWODB Task 1, achieves U-Recall 24.6 and lowest A-OSE 4148 among compared methods.
- Incremental VOC mAP surpasses existing baselines by 1.8-4.6% overall; calibration raises Task 2 U-Recall by 9.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feature orthogonalization (ablation) | Base model · M-OWODB Task 1 | U-Recall: 8.4 -> 18.2 | Helped |
| Add feature orthogonalization (ablation) | Base model · M-OWODB Task 1 | K-mAP: 59.8 -> 61.3 | Helped |
| Add prediction orthogonalization (ablation) | Feature-orthogonalized detector · M-OWODB Task 1 | U-Recall: +6.4% | Helped |
| Add feature and prediction orthogonalization (ablation) | Base model · M-OWODB Task 1 | A-OSE: 5922 -> 4148 | Helped |
| Add cross-task calibration (ablation) | Ours without calibration · M-OWODB Task 2 | U-Recall: +9.1% | Helped |
| Add cross-task calibration (ablation) | Ours without calibration · M-OWODB Task 2 | mAP (both): +1.1% | Helped |
| Use the full orthogonalization method instead of the base model | RandBox base model · PASCAL VOC 10 + 10 | mAP: 69.8 -> 72.3 | Helped |
| Use the full orthogonalization method instead of the base model | RandBox base model · PASCAL VOC 15 + 5 | mAP: 72.4 -> 74.7 | Helped |
| Use the full orthogonalization method instead of the base model | RandBox base model · PASCAL VOC 19 + 1 | mAP: 73.4 -> 75.6 | Helped |
Surprising findings
- Feature orthogonalization alone raises U-Recall substantially (8.4 to 18.2 on Task 1), and the paper notes the first-task mAP lags PROB/CAT-style baselines.
Limitations
- First-task known-class mAP lags some baselines (the paper concedes this and attributes it to later incremental learning).
- Unknown-class assignment relies on a heuristic maximum-softmax confidence criterion rather than a learned model.
Experimental setup
Datasets: COCO PASCAL VOC M-OWODB S-OWODB
Benchmarks: M-OWODB S-OWODB PASCAL VOC incremental splits
Metrics: U-Recall mAP wilderness impact absolute open-set error
Baselines: ORE OST OW-DETR UC-OWOD ALLOW PROB CAT RandBox ILOD Faster ILOD Meta-ILOD ROSETTA
Builds on: RandBox ORE PROB Fast R-CNN Gumbel-Max trick
Closest papers by meaning
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.92)
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.90)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.90)
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.88)
- Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.88)
- Towards 3D Objectness Learning in an Open World (NeurIPS 2025 · similarity 0.88)
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.