Research map / Detection and segmentation / CNN features
AirRoom: Objects Matter in Room Reidentification
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 AirRoom, a training-free coarse-to-fine room reidentification pipeline that uses multi-level object-oriented cues (global context, object patches, segmentation, keypoints) and releases four new Habitat-simulated room ReID datasets.
Problem
Room reidentification in cluttered indoor scenes, where global-descriptor and local-aggregation visual place recognition methods struggle to distinguish semantically similar rooms and handle viewpoint variations.
Approach
Use object-level information in a coarse-to-fine retrieval: global features pick top-5 rooms, then object and object-patch features (with patches built by Delaunay-triangulation-based receptive field expansion) and mutual-nearest-neighbor object-aware scoring pick top-2, and LightGlue keypoint matching selects the final room, with no training or fine-tuning.
Key results
- On HMReID, AirRoom reaches 93.80% accuracy versus 89.69% for AnyLoc, the strongest baseline in Table 1.
- AirRoom with NetVLAD features improves MPReID accuracy from 82.22% (Patch-NetVLAD group baseline is 64.32% at 4096 dims) to 89.38%, and ResNet50 on MPReID from 76.14% to 86.16% in the group-wise comparison.
- AirRoom consistently exceeds 85% across metrics for most global feature extractors in flexibility tests.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the patch score (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 66.68 -> 70.95 | Helped |
| add the object score (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 69.77 -> 70.95 | Helped |
| add Fine-Grained Retrieval (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 66.11 -> 70.95 | Helped |
| add both the patch score and object score (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 62.26 -> 70.95 | Helped |
| add the patch score and Fine-Grained Retrieval (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 59.39 -> 70.95 | Helped |
| add the object score and Fine-Grained Retrieval (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 63.44 -> 70.95 | Helped |
| add the global score (ablation) | AirRoom with ViT · HMReID | accuracy: 84.86 -> 86.58 | Helped |
| add the global score (ablation) | AirRoom with DINOv2 · HMReID | accuracy: 71.73 -> 70.95 | Hurt |
| replace Mask R-CNN with Semantic-SAM for instance segmentation (ablation) | AirRoom · HMReID | accuracy: 69.44 -> 70.95 | No clear effect |
| replace ResNet50 with DINOv2 as Object Feature Extractor (ablation) | AirRoom · HMReID | accuracy: 70.95 -> 68.67 | No clear effect |
| use max rather than mean for the object score (ablation) | AirRoom · HMReID | accuracy: 70.95 -> 71.02 | No clear effect |
| use mean rather than max for the patch score (ablation) | AirRoom · HMReID | accuracy: 71.02 -> 70.85 | No clear effect |
Surprising findings
- Removing the global score improves AirRoom-DINOv2 on HMReID (71.73% vs 70.95% accuracy), indicating the global prior can hurt when the global extractor is weak.
- Image retrieval (CVNet) and global-descriptor VPR baselines perform poorly on room ReID, with CVNet reaching only 17.45% accuracy on MPReID.
Limitations
- Not verified robustness to object rearrangements from movable objects, since the datasets lack such cases.
- Evaluation limited to four simulator-generated datasets with small numbers of rooms (19 to 105).
Experimental setup
Datasets: MPReID HMReID GibsonReID ReplicaReID
Benchmarks: MPReID HMReID GibsonReID ReplicaReID
Metrics: accuracy precision recall F1
Baselines: CVNet DINOv2 Patch-NetVLAD AnyLoc NetVLAD ResNet50
Builds on: Patch-NetVLAD AnyLoc LightGlue Mask R-CNN Semantic-SAM
Closest papers by meaning
- Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping (NeurIPS 2025 · similarity 0.86)
- Cubify Anything: Scaling Indoor 3D Object Detection (CVPR 2025 · similarity 0.85)
- UniDet3D: Multi-dataset Indoor 3D Object Detection (AAAI 2025 · similarity 0.84)
- Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection (CVPR 2025 · similarity 0.84)
- Seeing the Unseen: Visual Common Sense for Semantic Placement (CVPR 2024 · similarity 0.83)
- Physical AI Smart Spaces: A Large-Scale Benchmark for Multi-Camera 3D Perception in Smart Spaces (NeurIPS 2026 · similarity 0.83)
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.