UnLoc: Leveraging Depth Uncertainties for Floorplan Localization

1. Problem Statement

The paper addresses the challenge of sequential visual camera localization within indoor environments using 2D floorplans. While traditional methods rely on storage-intensive 3D models or extensive image databases, floorplans offer a lightweight, persistent, and robust alternative.

However, existing state-of-the-art methods, specifically F3Loc (Chen et al., 2024a), suffer from two critical limitations that hinder their practical deployment and generalization:

1. Lack of Uncertainty Modeling: F3Loc assumes all depth predictions are equally accurate. In real-world indoor scenes, depth estimation is often unreliable due to glass walls, open doorways, or featureless surfaces. Fusing these unreliable predictions without weighting them by confidence leads to erroneous pose estimates.
2. Dataset-Specific Depth Networks: F3Loc requires training a custom depth prediction network for every specific environment. This is impractical for rapid deployment in new or unseen spaces, as collecting sufficient depth data for retraining is costly and time-consuming.

2. Methodology: UnLoc

The authors propose UnLoc, an efficient, data-driven solution that integrates uncertainty estimation into monocular depth prediction and leverages pre-trained models. The pipeline consists of the following stages:

A. Input and Preprocessing

• Inputs: A sequence of RGB images, relative poses between frames, gravity direction, camera intrinsics, and a 2D occupancy grid floorplan.
• Gravity Alignment: Images are warped to align with the gravity vector (using roll and pitch angles) to ensure consistent orientation. A binary mask is generated to exclude invalid pixels caused by warping.

B. Feature Extraction with Off-the-Shelf Models

• Instead of training custom encoders, UnLoc utilizes pre-trained monocular depth models (specifically Depth Anything v2 fine-tuned for indoor scenes).
• The model extracts features from the gravity-aligned image, which are then processed by a Masked Attention Mechanism. This mechanism focuses attention on observable regions (using the gravity mask) to predict depth.

C. Uncertainty-Aware Depth Prediction

• The core innovation is modeling depth predictions as explicit probability distributions rather than single point estimates.
• The network outputs two 1D vectors:
  1. $\hat{d}_t$: Predicted floorplan depth.
  2. $\hat{b}_t$: Scale parameter representing aleatoric uncertainty (data noise).
• Distribution Model: The depth is modeled as a Laplace distribution centered at $\hat{d}_t$ with scale $\hat{b}_t$. The Laplace distribution is chosen for its heavier tails (robustness to outliers in challenging scenes) and computational efficiency for closed-form likelihood calculations.
• Training Loss: The model is trained using the negative log-likelihood of the Laplace distribution, encouraging accurate depth while penalizing over-confident predictions in ambiguous regions.

D. Uncertainty-Aware Matching

• The predicted depth and uncertainty are used to compute an observation likelihood over the entire floorplan.
• The likelihood $p(o_t | s_t)$ is calculated as the product of independent Laplace distributions for rays cast from the camera pose $s_t$.
• Key Mechanism: When uncertainty ($\hat{b}_t$) is high (e.g., near glass or doors), the distribution flattens, naturally down-weighting the contribution of that observation to the pose estimation. This prevents unreliable data from corrupting the localization.

E. Sequential Fusion via Histogram Filter

• A Histogram Filter (Bayesian filtering) fuses the observation likelihoods over time.
• The posterior probability $p(s_t | o_t, t_t)$ is updated by combining the current observation likelihood with the prior belief propagated through a motion model (ego-motion).
• The uncertainty directly influences the update step, allowing the filter to rely more heavily on confident predictions.

F. Post-Processing Optimization

• To correct residual drift, a lightweight SE(2) optimization is performed on the last $k$ frames (typically $k=10$).
• This step minimizes an uncertainty-weighted L1 loss between predicted and floorplan depths, refining the trajectory with a global rigid transformation (rotation and translation).

3. Key Contributions

1. Uncertainty-Aware Depth Prediction: The first method to model floorplan depth as an explicit probability distribution (Laplace) for localization. This allows the system to dynamically weight observations based on reliability, significantly improving robustness in ambiguous indoor environments.
2. Leveraging Off-the-Shelf Models: UnLoc eliminates the need for per-environment training by utilizing state-of-the-art pre-trained monocular depth models (Depth Anything v2). This enables zero-shot generalization to unseen spaces.
3. Probabilistic Fusion: The integration of uncertainty into the histogram filter creates a principled framework for sequential localization that outperforms deterministic approaches.
4. Efficiency: The method maintains real-time performance while achieving superior accuracy, making it suitable for practical AR and robotics applications.

4. Experimental Results

The authors evaluated UnLoc on Gibson (synthetic) and LaMAR HGE/CAB (real-world) datasets.

• Performance on Long Sequences (100 frames):
  • On the challenging LaMAR HGE dataset, UnLoc achieved 100% success rate (within 1m error), compared to F3Loc's 36.4%. This represents a 2.7x improvement in recall.
• Performance on Short Sequences (15 frames):
  • UnLoc achieved a 50.6% success rate, whereas F3Loc only achieved 1.2%. This is a 42.2x improvement.
  • This is critical for applications requiring rapid localization (e.g., "I just walked in, where am I?").
• Generalization:
  • Models trained on LaMAR HGE successfully localized in the unseen LaMAR CAB building without retraining, whereas F3Loc failed completely.
• Ablation Studies:
  • Replacing F3Loc's custom encoder with Depth Anything v2 improved results, but adding uncertainty modeling provided the most significant gains.
  • The Laplace distribution model outperformed a Gaussian model, particularly on shorter sequences, due to better handling of outliers.
• Runtime:
  • UnLoc processes frames in approximately 0.18s (feature extraction) + 0.88s (matching) on an NVIDIA Quadro RTX 6000. While slightly slower than F3Loc's custom network, it remains practical for many applications, especially given the massive accuracy gains.

5. Significance

UnLoc represents a paradigm shift in visual floorplan localization. By moving away from the assumption of "perfect" depth estimates and the requirement for environment-specific training, it offers a scalable, robust, and generalizable solution.

• Practical Impact: It enables devices to localize in new buildings immediately upon arrival using only a floorplan and a standard camera, without the need for prior 3D scanning or extensive data collection.
• Robustness: The explicit handling of uncertainty makes the system resilient to the "messy" reality of indoor environments (glass, occlusions, lighting changes), which has historically been a failure point for geometric localization methods.
• Future Direction: The paper establishes that pre-trained foundation models for depth, when combined with probabilistic reasoning, can solve complex localization tasks more effectively than specialized, trained-from-scratch pipelines.

The code and materials are publicly available at: https://github.com/matthias-wueest/UnLoc.