DisCo-FLoc: Using Dual-Level Visual-Geometric Contrasts to Disambiguate Depth-Aware Visual Floorplan Localization

TL;DR

DisCo-FLoc tackles depth-aware visual floorplan localization by remedying multimodal ambiguity without semantic labels. It introduces a depth-aware Ray Regression Predictor (RRP) to generate a Depth-Aware FLoc Probabilistic Map (DAFPM) of candidate poses, followed by a dual-level visual-geometric contrastive learning (CL) module that yields a Disambiguation Probability Map (DPM) by aligning depth-aware visual features with localized floorplan geometry. Positive pairs strictly couple an image to the corresponding floorplan crop at a fixed pose, while negatives sample across positions and orientations to enforce robust cross-modal matching via a PointInfoNCE loss, with the depth encoder frozen to preserve depth estimation expertise. Experimental results on Gibson and Structured3D(full) show clear SOTA gains over both semantics-free and semantics-enabled baselines, underscoring the effectiveness of depth-aware priors and disambiguation in reducing pose ambiguity and improving localization accuracy. The approach significantly narrows the gap between purely positional and orientation-aware localization without relying on semantic annotations, enabling more scalable indoor localization in real-world deployments.

Abstract

Since floorplan data is readily available, long-term persistent, and robust to changes in visual appearance, visual Floorplan Localization (FLoc) has garnered significant attention. Existing methods either ingeniously match geometric priors or utilize sparse semantics to reduce FLoc uncertainty. However, they still suffer from ambiguous FLoc caused by repetitive structures within minimalist floorplans. Moreover, expensive but limited semantic annotations restrict their applicability. To address these issues, we propose DisCo-FLoc, which utilizes dual-level visual-geometric Contrasts to Disambiguate depth-aware visual Floc, without requiring additional semantic labels. Our solution begins with a ray regression predictor tailored for ray-casting-based FLoc, predicting a series of FLoc candidates using depth estimation expertise. In addition, a novel contrastive learning method with position-level and orientation-level constraints is proposed to strictly match depth-aware visual features with the corresponding geometric structures in the floorplan. Such matches can effectively eliminate FLoc ambiguity and select the optimal imaging pose from FLoc candidates. Exhaustive comparative studies on two standard visual Floc benchmarks demonstrate that our method outperforms the state-of-the-art semantic-based method, achieving significant improvements in both robustness and accuracy.

Figures (4)

• Figure 1: (b) and (c) respectively show the probability maps generated by SemRayLoc and our RRP for FLoc based on the monocular RGB image in (a). Both of them suffer from ambiguous localization caused by repetitive structures. (d) shows our visual-geometric CL-based DPM. (e) shows the final FLoc by using our DPM to disambiguate the probability map in (c). (f) shows that our method significantly outperforms existing SOTA methods across multiple localization accuracies on the challenging Structured3D(full) dataset.
• Figure 2: An illustration of the depth-aware RRP, which maps depth estimation expertise to probabilities across different bins in Eq. (\ref{['eq2']}).
• Figure 3: (a) shows the collection of positive and negative samples for visual-geometric CL. (b) shows the visual-geometric CL (training) or contrastive disambiguation (inference) used in (c). The contrastive loss $\mathcal{L}_{CL}$ is only used during training. The CLS token, which contains global depth information, is fused with the depth tokens to form a query $\mathcal{Q}$. (c) shows the FLoc disambiguation. The depth-aware RRP shown in Fig. \ref{['fig2']} is first used to generate a DAFPM. The Top-$X$ poses with the highest probabilities in DAFPM are selected as FLoc candidates, and their corresponding floorplan structures are cropped according to a specific size. The frozen DINO V2 and pre-trained ResNet18 are used to encode features for computing similarities between the visual image and the floorplan structures during contrastive disambiguation, yielding a DPM. The DPM is fused with the DAFPM using a weight $w$ to select the optimal FLoc from the candidates.
• Figure 4: The impact of different numbers $X$ of candidates on FLoc performance. To balance computational cost and performance, we selected $X$ = 100 in our experiments.