Research map / Detection and segmentation / Open-set and few-shot detection
Percept, Memory, and Imagine: World Feature Simulating for Open-Domain Unknown 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 World Feature Simulation (WFS), which synthesizes virtual unknown features from multi-level perception (local, Mamba global, graph connective), a memory codebook, and an imagination bank to detect unknown objects under combined covariate and semantic shifts.
Problem
Detecting unknown out-of-distribution objects in unseen target domains with both covariate shifts (style changes) and semantic shifts, without any auxiliary data (Open-Domain Unknown Object Detection, ODU-OD).
Approach
Mimics human perception, memory, and imagination: multi-level features (local, Mamba-global, graph-connective) are fused with a learnable memory codebook; a modulated memory yields an imagination bank, and a recurrent unknown-feature generator synthesizes outlier features for training an uncertainty-based unknown detector.
Key results
- On ODU-OD with VOS as base, VOS+WFS reduces FPR95 from about 88.7% to 80.2% on MS-COCO Acrylic and raises AUROC to 74.31%.
- On unsupervised OOD-OD with BDD source, SAFE+WFS reaches FPR95 21.84 / AUROC 93.41 versus SAFE baseline 32.56 / 80.30 (VOC source reported separately).
- On OSOD VOC-COCO-20, OpenDet+Ours improves APU from 14.93 to 17.19 and reduces WI from 14.95 to 11.98.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add WFS to VOS | VOS · OOD-OD, VOC source, MS-COCO OOD | FPR95: 47.53 -> 37.91 | Helped |
| add WFS to VOS | VOS · OOD-OD, VOC source, OpenImages OOD | AUROC: 85.23 -> 89.78 | Helped |
| add WFS to VOS | VOS · OOD-OD, BDD source, MS-COCO OOD | FPR95: 44.27 -> 28.66 | Helped |
| add WFS to VOS | VOS · OOD-OD, BDD source, OpenImages OOD | AUROC: 88.52 -> 94.13 | Helped |
| add WFS to Clip-gap | Clip-gap · Single-DGOD Night-Sunny | mAP: 36.9 -> 38.5 | Helped |
| add WFS to OpenDet | OpenDet · OSOD VOC | mAPK: 80.02 -> 80.53 | Helped |
| add the memory recorder to perception and imagination generator (ablation) | WFS · ODU-OD, Acrylic, VOC source | FPR95: 81.07% -> 80.16% | Helped |
| add the imagination-based unknown-feature generator to perception and memory (ablation) | WFS · ODU-OD, Acrylic, VOC source | FPR95: 82.85% -> 80.16% | Helped |
| combine local, global, and connective perception (ablation) | WFS · ODU-OD, Acrylic, VOC source | FPR95: 80.58% -> 80.16% | Helped |
| use 16 codewords instead of the reported setting (ablation) | WFS memory recorder · ODU-OD, Acrylic, VOC source | FPR95: 80.54% | Hurt |
| use 32 codewords instead of the reported setting (ablation) | WFS memory recorder · ODU-OD, Acrylic, VOC source | FPR95: 80.33% | Hurt |
Limitations
- Constructed Open-Domain dataset is derived from synthetic style transfer rather than real adverse-domain imagery
- Uses a fixed threshold of 0.95 on the uncertainty score that may need tuning
- Hyperparameters (alpha, lambda, Q) chosen empirically; higher Q increases overfitting
Experimental setup
Datasets: PASCAL VOC BDD-100k MS-COCO OpenImages
Benchmarks: Open-Domain Unknown Object Detection dataset (ODU-OD) OOD-OD (VOC vs MS-COCO/OpenImages, BDD) OSOD VOC-COCO-20/40/60 Single-DGOD Diverse Weather
Metrics: FPR95 AUROC mAP WI AOSE APU
Baselines: SR-VAE DFDD VOS SAFE SIREN-KNN OpenDet S-DGOD Clip-gap MSP ODIN Energy score
Builds on: VOS Faster R-CNN Vision Mamba SAFE CLIPstyler
Closest papers by meaning
- Knowing the Unknown: Interpretable Open-World Object Detection via Concept Decomposition Model (ICML 2026 · similarity 0.88)
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.87)
- Exploring Orthogonality in Open World Object Detection (CVPR 2024 · similarity 0.87)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.86)
- Looking Beyond the Known: Towards a Data Discovery Guided Open-World Object Detection (NeurIPS 2025 · similarity 0.86)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.86)
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.