💻 computer science

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

DisCo-FLoc addresses the ambiguity and annotation limitations in visual floorplan localization by introducing a ray regression predictor and a novel dual-level contrastive learning framework that matches depth-aware visual features with geometric structures without requiring additional semantic labels.

Original authors: Shiyong Meng, Tao Zou, Bolei Chen, Chaoxu Mu, Jianxin Wang

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shiyong Meng, Tao Zou, Bolei Chen, Chaoxu Mu, Jianxin Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a massive, modern office building. The walls are all white, the floors are all gray, and every hallway looks exactly like the next. You pull out your phone to ask, "Where am I?"

If you rely on a standard GPS, it won't work inside. If you rely on a map that just shows the shape of the rooms (a floorplan), it's still tricky because the map doesn't show the coffee machine, the potted plant, or the specific angle you're standing at. You might look at a corner and think, "That's the breakroom," but it could actually be the conference room down the hall because they look identical. This is the problem of Visual Floorplan Localization: figuring out exactly where a camera is inside a building just by looking at a photo and a simple line drawing of the floor.

The paper you shared introduces a new solution called DisCo-FLoc. Here is how it works, broken down into simple concepts and analogies.

The Problem: The "Hall of Mirrors" Effect

Existing methods try to match the photo you take with the floorplan. They do this by shooting invisible "lasers" (rays) from your camera position to the walls to measure distances.

The Issue: In a building with repetitive rooms (like a hotel with 50 identical rooms), the lasers might hit a wall 5 meters away in Room A, and also 5 meters away in Room B. The computer gets confused. It thinks, "I could be in Room A or Room B." This is called ambiguity.
The Old Fix: Some previous methods tried to use "semantic labels" (telling the computer, "That's a door," "That's a window"). But getting a human to label every single door and window on every floorplan is expensive, slow, and often impossible.

The Solution: DisCo-FLoc

The authors propose a two-step process that acts like a smart detective, solving the mystery without needing a human to label the doors.

Step 1: The "Depth-Savvy" Guess (The Ray Regression Predictor)

First, the system uses a pre-trained AI expert in depth estimation (an AI that is really good at judging how far away things are in a photo).

The Analogy: Imagine you are blindfolded but have a super-sensitive sense of touch. You reach out and feel the distance to the nearest wall in front of you, to your left, and to your right.
What it does: The AI looks at your photo and predicts a series of "distance guesses." It creates a Probability Map. Instead of saying "You are here," it says, "You are probably in these 100 spots, but I'm not 100% sure yet."
The Result: It generates a list of top candidates (e.g., "Maybe you are in the hallway, maybe you are in the lobby").

Step 2: The "Double-Check" (Visual-Geometric Contrast)

This is the magic part. The system now has a list of 100 guesses. It needs to pick the one correct answer. It does this using a technique called Contrastive Learning.

The Analogy: Imagine you are trying to find your lost keys in a messy room. You have a list of 100 places they might be.
- The "Positive" Match: You look at the spot where you think they are and compare it to a mental image of what your keys look like in that specific spot. "Yes, that matches!"
- The "Negative" Matches (The Tricky Part): To be sure, you also look at spots that are similar but wrong.
  - Position Check: "Is this the exact spot, or is it the spot 3 feet to the left?" (This helps distinguish between two identical-looking corners).
  - Orientation Check: "Am I facing North, or am I facing South?" (This helps distinguish between a room that looks the same from two different angles).
How DisCo-FLoc does it: It takes the photo and the floorplan. It creates "fake" wrong answers by shifting the location slightly or rotating the angle. It then teaches the AI: "This photo matches the floorplan here, but it definitely does not match the floorplan there or facing that way."
The Result: By learning what doesn't fit, the AI becomes incredibly good at spotting what does fit. It eliminates the "maybe" spots and picks the single best location.

Why is this a big deal?

No Labels Needed: It doesn't need a human to draw boxes around doors or windows. It learns purely from the geometry (the shapes and lines) of the floorplan.
Beats the Experts: The paper shows that this method works better than current state-of-the-art methods, even those that do use expensive labels.
Direction Matters: The system is surprisingly good at knowing not just where you are, but which way you are facing. It realized that knowing the direction (North vs. South) is often the key to solving the puzzle.

Summary

Think of DisCo-FLoc as a navigation system that doesn't need a detailed map with street names. Instead, it uses a super-accurate sense of distance to make a list of possibilities, and then uses a "process of elimination" (comparing the right spot against slightly wrong spots) to find the one true location. It's like solving a maze by realizing that while many paths look the same, only one path fits the specific angle you are walking in.

1. Problem Statement

Visual Floorplan Localization (FLoc) aims to determine the 2D camera position $(x, y)$ and orientation $\theta$ within a building based on a monocular RGB image and a corresponding 2D floorplan. While floorplans are robust, persistent, and readily available, existing methods face two critical challenges:

Ambiguity from Repetitive Structures: Minimalist floorplans often contain repetitive geometric patterns (e.g., identical rooms or corridors). Traditional geometric matching methods (like ray-casting) struggle to distinguish between these similar structures, leading to multimodal probability distributions and localization errors.
Limitations of Semantic Reliance: Some state-of-the-art (SOTA) methods attempt to resolve ambiguity by using semantic labels (e.g., "door," "window," "room type"). However, these labels are often unavailable in standard floorplans or require expensive manual annotation, limiting the scalability and applicability of such methods.

The core problem is how to disambiguate visual localization in repetitive environments without relying on additional semantic annotations.

2. Methodology: DisCo-FLoc

The authors propose a hierarchical, two-stage framework called DisCo-FLoc (Disambiguation via Contrastive Learning for FLoc).

Stage 1: Depth-Aware Ray Regression Predictor (RRP)

Instead of using generic image classifiers, the method treats 2D ray-casting as a specialized depth estimation task.

Architecture: It utilizes a depth-aware encoder (specifically DINO V2 from Depth Anything V2) pre-trained on dense monocular depth estimation tasks. This allows the model to leverage expert depth features without training from scratch.
Mechanism: The encoder extracts visual features, which are processed through a convolutional layer and a vertical weighted summation to form queries. These queries predict a probability distribution of distances to the nearest wall along specific rays.
Output: The RRP generates a Depth-Aware FLoc Probabilistic Map (DAFPM). This map provides a set of high-probability candidate poses but still suffers from ambiguity due to repetitive structures.

Stage 2: Visual-Geometric Contrastive Learning (CL) for Disambiguation

To resolve the ambiguity in the DAFPM, a novel contrastive learning framework is introduced to strictly match visual features with geometric structures.

Dual-Level Constraints: The method employs two types of negative samples to enforce strict matching:
1. Position-Level Negatives: Samples where the observation position is changed (either within the same floorplan or across different floorplans). This forces the model to distinguish the correct geometry from other similar structures in different locations.
2. Orientation-Level Negatives: Samples where the observation orientation is rotated. This enhances the model's sensitivity to direction, distinguishing the correct pose from incorrect orientations.
Training Process: Positive pairs consist of an RGB image and its corresponding floorplan crop (constrained by a unique pose). Negative pairs are constructed by varying positions and orientations. A PointInfoNCE loss is used to maximize the similarity of positive pairs while minimizing similarity with negatives.
Disambiguation Inference:
1. The RRP generates the top- $X$ candidate poses (e.g., $X=100$ ) from the DAFPM.
2. For each candidate, the corresponding local floorplan structure is cropped.
3. A frozen depth-aware encoder and a pre-trained ResNet-18 encode the visual image and floorplan crops, respectively.
4. Cosine similarity scores generate a Disambiguation Probability Map (DPM).
5. The final pose is selected by fusing the DPM and DAFPM with a weighted sum, choosing the candidate with the highest total score.

3. Key Contributions

Hierarchical Framework: A novel pipeline that first generates candidates using a depth-expertise-enhanced RRP and then disambiguates them using contrastive learning, avoiding the need for semantic labels.
Dual-Level Contrastive Learning: The introduction of position-level and orientation-level negative samples to strictly constrain the matching between visual features and geometric structures, effectively eliminating false correlations caused by repetitive layouts.
Semantic-Free SOTA Performance: The method achieves superior performance compared to existing SOTA methods, including those that utilize semantic annotations (e.g., SemRayLoc), proving that geometric contrasts alone can resolve localization ambiguity.
Depth-Aware Encoder Integration: Successfully adapting pre-trained depth estimation models (DINO V2) for visual floorplan localization, bridging the gap between depth estimation and 2D ray-casting.

4. Experimental Results

The method was evaluated on two standard benchmarks: Gibson (general and forward motions) and Structured3D(full) (challenging, photorealistic, fully furnished environments).

Performance on Gibson: DisCo-FLoc outperformed the strong baseline "3DP & RSK" (which combines geometric priors and unsupervised room-style learning) by significant margins (e.g., +8.8% improvement in R@1m on Gibson(f)).
Performance on Structured3D(full): This dataset is more challenging due to furniture and occlusion. DisCo-FLoc achieved 67.0% recall at 1m accuracy, significantly outperforming:
- The best non-semantic baseline (3DP & RSK) by 12.3%.
- The best semantic-based baseline (SemRayLocs + 3DP & RSK) by 5.5%.
- Notably, it narrowed the performance gap between positional accuracy (R@1m) and directional accuracy (R@1m 30°), indicating high precision in orientation estimation without semantic cues.
Ablation Studies:
- Both position-level and orientation-level negative samples contributed to performance, with orientation-level constraints proving more critical for disambiguation.
- The use of the CLS token (global depth features) and pose perturbations during training significantly boosted robustness.
- Optimal parameters were found at a candidate count $X=100$ and a fusion weight $w=0.5$ .

5. Significance

Practical Applicability: By eliminating the dependency on semantic annotations, DisCo-FLoc can be deployed in real-world scenarios where floorplans are available but lack detailed semantic labeling (e.g., legacy building maps).
Robustness: The method demonstrates superior robustness to repetitive structures and visual occlusions compared to purely geometric or semantic-dependent methods.
Paradigm Shift: It establishes that visual-geometric contrastive learning is a powerful alternative to semantic supervision for resolving localization ambiguities, offering a new direction for visual localization in indoor environments.
Efficiency: The two-stage approach allows for parallel feature extraction during the disambiguation phase, maintaining real-time capabilities despite the complex matching process.