Research map / Detection and segmentation / Open-set and few-shot detection
Semi-supervised 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.
Introduces semi-supervised open-world object detection (SS-OWOD) and SS-OWFormer, a Deformable DETR-based detector that uses feature alignment on augmented images and object-query-guided pseudo-labeling for unknown objects, evaluated on COCO, PASCAL VOC, Objects365 and DOTA.
Problem
Open-world object detection relies on fully annotated data from a human oracle during incremental learning stages, which is costly, and existing pseudo-labeling heuristics for unknowns (channel averaging of backbone features) are suboptimal.
Approach
Casts OWOD incremental stages as semi-supervised learning: a frozen detached copy of the previous model and a mapping network align object query embeddings between original and augmented images, while unknown pseudo-labels come from decoder object queries modulated with multi-scale encoder features and pooled at predicted boxes.
Key results
- On COCO, SS-OWFormer with 10% labeled data outperforms fully supervised OW-DETR on mAP across tasks; SS-OWFormer (Task 1) reaches 12.26 U-Recall vs 7.5 for OW-DETR.
- SS-OWFormer (50%) achieves 11.13 U-Recall and 55.51 mAP on Task 1 of the satellite split, versus OW-DETR's 7.33 and 52.78 (in the table with 50% labels).
- On PASCAL VOC 10+10, SS-OWFormer (50%) reaches 65.2 mAP, comparable to OW-DETR's 65.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use object query-guided pseudo-labeling instead of baseline feature averaging (ablation) | SS-OWFormer · MS COCO Task-1 | U-Recall: 6.17 -> 12.26 | Helped |
| use encoder features instead of backbone features for pseudo-labeling (ablation) | SS-OWFormer · MS COCO Task-1 | U-Recall: 6.17 -> 8.06 | Helped |
| add decoder object queries to encoder-feature pseudo-labeling (ablation) | SS-OWFormer · MS COCO Task-1 | U-Recall: 8.06 -> 12.26 | Helped |
| use maximum instead of mean for objectness scores (ablation) | SS-OWFormer · MS COCO Task-1 | U-Recall: 12.26 -> 11.25 | Hurt |
| use semi-supervised learning with 50% labeled data instead of no SSL (ablation) | SS-OWFormer · Satellite OWOD Split-1 Task-2 | mAP Overall: 63 -> 68.06 | Helped |
| use semi-supervised learning with 75% labeled data instead of no SSL (ablation) | SS-OWFormer · Satellite OWOD Split-1 Task-2 | mAP Overall: 65.2 -> 69.04 | Helped |
| use color jitter, Gaussian blur, and greyscale together (ablation) | SS-OWFormer · MS COCO Task-2 (50% labeled) | U-Recall: 9.57 -> 10.03 | Helped |
| add posterizing and solarizing to color jitter, Gaussian blur, and greyscale (ablation) | SS-OWFormer · MS COCO Task-2 (50% labeled) | mAP: 38.71 -> 38.13 | Hurt |
| use semi-supervised incremental learning with 50% labeled data | SS-OWFormer · Satellite OWOD-S Task-1 | U-Recall: 2.5 -> 7.6 | Helped |
| use SS-OWFormer instead of OW-DETR | MS COCO Task-1 · MS COCO Task-1 | U-Recall: 7.5 -> 12.26 | Helped |
| use SS-OWFormer instead of OW-DETR | PASCAL VOC 10+10 · PASCAL VOC 10+10 | mAP: 65.7 -> 65.2 | Hurt |
| use SS-OWFormer instead of OW-DETR | Objects365 · Objects365 Task-1 | U-Recall: 13.6 -> 16.7 | Helped |
Surprising findings
- The state-of-the-art OWOD detector's performance dramatically deteriorates in the semi-supervised setting, per the abstract.
- Taking the maximum instead of the mean of objectness scores for pseudo-labeling slightly lowers unknown recall (11.25 vs 12.26).
Limitations
- No explicit mechanism to forget previously learned categories.
- Performance on satellite imagery remains relatively low due to OWOD complexity and partial annotation.
- Uses a fixed top-k pseudo-label selection and a fixed query count tuned for satellite images.
Experimental setup
Datasets: MS COCO PASCAL VOC DOTA Objects365
Benchmarks: MS COCO OWOD split PASCAL VOC 10+10 DOTA OWOD-S splits Objects365 OWOD splits
Metrics: mAP unknown recall (U-Recall)
Baselines: OW-DETR ORE ORE-EBUI Faster R-CNN
Builds on: OW-DETR Deformable DETR ORE FixMatch
Closest papers by meaning
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.89)
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.89)
- OW-VAP: Visual Attribute Parsing for Open World Object Detection (ICML 2025 · similarity 0.88)
- Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.88)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.88)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.87)
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.