Abstract
Feature extraction plays an important role in visual localization. Unreliable features on dynamic objects or repetitive regions will interfere with feature matching and challenge indoor localization greatly. To address the problem, we propose a novel network, RaP-Net, to simultaneously predict region-wise invariability and point-wise reliability, and then extract features by considering both of them. We also introduce a new dataset, named OpenLORIS-Location, to train the proposed network. The dataset contains 1553 images from 93 indoor locations. Various appearance changes between images of the same location are included and can help the model to learn the invariability in typical indoor scenes. Experimental results show that the proposed RaP-Net trained with OpenLORIS-Location dataset achieves excellent performance in the feature matching task and significantly outperforms state-of-the-arts feature algorithms in indoor localization.
Introduction
Visual localization—the ability of a robot or device to determine its position based on an image—is notoriously difficult indoors. Unlike outdoor environments with distinct landmarks, indoor spaces are filled with "unreliable" features. Dynamic objects, such as pedestrians or moved furniture, create features that are transient and non-repeatable. Furthermore, repetitive textures like visually identical pillars, ceiling tiles, or carpet patterns can lead to "aliasing," where a system incorrectly identifies its location.
Existing deep-learning methods typically focus on point-wise reliability, identifying distinct pixels like corners or edges. However, they often lack the global context needed to determine if a region is a stable part of the environment. RaP-Net addresses this by scoring image features based on both local reliability (point-wise) and environmental stability (region-wise).
Methods
The proposed RaP-Net (Region-wise and Point-wise weighting Network) utilizes a dual-weighting strategy to extract robust features:
- Shared Backbone: A truncated VGG-16 architecture is used to extract dense feature maps.
- Point-wise Weight Estimator: This component identifies pixels that are repeatable and distinct across different viewpoints, functioning as a standard feature detector.
- Region-wise Weight Estimator: This module uses an attention mechanism with three parallel branches (3x3, 5x5, and 7x7 kernels) to capture multi-scale context. This allows the network to evaluate larger areas and distinguish between invariable (static) and dynamic regions.
- Feature Fusion: During inference, the final feature score combines both weights. Features in stable regions are prioritized, while those on dynamic objects are suppressed.
- Training Strategy: The model is trained using metric learning with a triplet loss, encouraging the production of similar global descriptors for the same location despite appearance changes.
The OpenLORIS-Location Dataset
To train the region-wise attention mechanism, the authors introduced the OpenLORIS-Location dataset. It contains 1,553 images from 93 indoor locations, capturing significant appearance changes such as lighting variations, moved objects, and human activity.
Unlike existing attention-based datasets that focus on outdoor landmarks, OpenLORIS-Location provides the necessary data to teach networks what "stable" looks like in mundane office or corridor environments.
Results and Discussion
RaP-Net was evaluated against state-of-the-art methods like SuperPoint, D2-Net, and HF-Net across several benchmarks:
- Feature Matching: On the HPatches dataset, RaP-Net achieved higher Mean Matching Accuracy (MMA) than competitors, particularly under strict error thresholds (1-3 pixels).
- Visual SLAM Performance: In the OpenLORIS-Scene "office" sequence, RaP-Net achieved a 100% Correct Rate (CR) for tracking, significantly outperforming SuperPoint (67.9%) and HF-Net (98.4%).
- Robustness to Dynamics: Qualitative heatmaps demonstrate that RaP-Net successfully ignores pedestrians and focuses on stable structures like ceilings and permanent walls, whereas previous methods like D2-Net often triggered on moving people.
The key insight is that context is essential for localization. A pixel might be a distinct "corner," but if it resides on a person's backpack, it is useless for long-term mapping. RaP-Net’s region-wise weighting provides the semantic context to identify and suppress such unreliable regions.
Conclusion
RaP-Net demonstrates that combining point-wise reliability with region-wise invariability significantly improves indoor localization robustness. Future work will focus on optimizing the architecture for real-time performance on mobile robots and expanding the OpenLORIS-Location dataset to include a wider variety of indoor environments like warehouses and hospitals.
Summary updated