DisCo-FLoc: Visual Floorplan Localization

• DisCo-FLoc is a monocular visual floorplan localization system that predicts laser-style depth rays and leverages dual-level contrastive learning to disambiguate symmetric layouts.
• The framework combines a ray regression predictor with a ResNet-18 based encoder to perform robust geometry and orientation-level discrimination without semantic supervision.
• Empirical evaluations on benchmarks such as Gibson and Structured3D demonstrate significant improvements in recall and pose accuracy over state-of-the-art methods.

DisCo-FLoc is a monocular visual floorplan localization framework that introduces dual-level visual–geometric contrastive learning for robust, depth-aware localization, disambiguating repetitive or symmetric layouts without requiring semantic supervision. It addresses key challenges in floorplan localization (FLoc)—notably structural ambiguity and reliance on expensive annotation—by combining a specialized ray regression predictor with a fine-grained, multi-level contrastive disambiguation mechanism (Meng et al., 5 Jan 2026).

1. Pipeline Overview

DisCo-FLoc operates as a two-stage pipeline for single RGB image-based localization within a known floorplan geometry:

• Stage 1: Ray Regression Predictor (RRP)
  • Given an input image $\mathcal{I}$, the RRP predicts $N$ discrete “laser-style” distances $\vec{d} = (d_1, \dots, d_N)$ along fixed orientations to the nearest structural boundary.
  • The backbone is a frozen DINO V2 depth encoder. Features are reduced to dimension $D$ via $1 \times 1$ convolution, and depth-ray queries $q_i \in \mathbb{R}^D$ are aggregated vertically per image column.
  • A multi-head attention layer predicts, for each ray $i$, a probability distribution $P_{i,1\dots D}$ over $D$ depth bins.
  • Depth regression: For each ray,

  $$d_i = \sum_{k=1}^D P_{i,k} d_k, \qquad d_k = \bigl(d_{\min}^\gamma + \tfrac{k}{D}(d_{\max}^\gamma - d_{\min}^\gamma)\bigr)^{1/\gamma} \tag{1}$$ - Training loss: $N$0 plus a cosine “shape” term:

  $N$1 - Depth-Aware Floorplan Probabilistic Map (DAFPM): At inference, candidate $N$2 poses are tiled, ground-truth rays are rendered per pose, and compared to the predictions to yield $N$3. The $N$4 most likely candidates progress to the next stage.

• Stage 2: Visual–Geometric Contrastive Disambiguation

  • Each candidate $N$5 is associated with a 5 m $N$6 5 m floorplan patch, aligned to $N$7.
  • A contrastive learning mechanism is used to score fine-grained consistency between visual depth features and geometric layout in the patch.

This sequence enables DisCo-FLoc to first generate plausible localization candidates using a geometric prior and subsequently resolve ambiguities arising from floorplan regularities via contrastive matching.

2. Dual-Level Visual–Geometric Contrastive Learning

To discriminate between visually and geometrically similar locations, DisCo-FLoc pre-trains a ResNet-18 floorplan encoder with dual contrastive constraints while freezing the visual depth backbone.

• Position-Level Contrast:
  • Positive pairs: image $N$8 and its corresponding floorplan patch $N$9 at $\vec{d} = (d_1, \dots, d_N)$0, plus small spatial ($\vec{d} = (d_1, \dots, d_N)$1 m) and orientation ($\vec{d} = (d_1, \dots, d_N)$2 rad) perturbations.
  • Negative pairs are drawn from:
  • Inner-floorplan: patches 1.5–3 m away within the same layout.
  • Cross-floorplan: patches from other floorplans.
• Orientation-Level Contrast:
  • Negative pairs: same $\vec{d} = (d_1, \dots, d_N)$3 but heading $\vec{d} = (d_1, \dots, d_N)$4.
• PointInfoNCE Loss:
  • With negative set $\vec{d} = (d_1, \dots, d_N)$5 (for position and orientation levels):

  $\vec{d} = (d_1, \dots, d_N)$6

  $\vec{d} = (d_1, \dots, d_N)$7

  where $\vec{d} = (d_1, \dots, d_N)$8 is a temperature and “$\vec{d} = (d_1, \dots, d_N)$9” denotes cosine similarity.

• Training Objective:

$D$0

In practice, a single $D$1 is used as both negative types are included.

The dual-level scheme addresses both translational and rotational ambiguities, crucial for distinguishing between symmetric or repetitive structures.

3. Feature Extraction and Fusion

The feature encoders and inference logic operate as follows:

• Visual Feature Extraction:

  • The DINO V2 depth encoder is frozen for visual input.
  • A CLS (classification) token representing global depth is concatenated to the vertical column-token aggregation for the image.
• Floorplan Feature Extraction:
  • ResNet-18 is trained using $D$2 for geometric layout encoding.
• Candidate Matching and Scoring:
  • For each candidate $D$3, vision feature $D$4 and floorplan feature $D$5 are compared:

  $D$6

  The collection $D$7 is normalized across the $D$8 candidates using SoftMax to yield a disambiguation probability $D$9.

• Final Pose Selection:

$1 \times 1$0

where $1 \times 1$1 (set to 0.5). The candidate with the highest $1 \times 1$2 is selected as the localization output.

4. Empirical Evaluation

DisCo-FLoc is evaluated on Gibson(f), Gibson(g), and Structured3D(full), with metrics including Recall@0.1 m, 0.5 m, 1 m, and 1 m within $1 \times 1$3.

Method	[email protected] m	[email protected] m	R@1 m	R@1 m 30° (Gibson(f))
PF-net	0.0	2.0	6.9	1.2
MCL	1.6	4.9	12.1	8.2
LASER	0.4	6.7	13.0	10.4
F³Loc	4.7	28.6	36.6	35.1
3DP	5.3	33.2	39.8	38.4
RSK	8.3	38.5	45.3	43.6
Ours w/o Dis.	12.0	45.8	50.6	49.2
DisCo-FLoc	13.1	50.9	56.7	55.4

• On Structured3D(full), DisCo-FLoc achieves 10.0/59.0/67.0/66.0 at [email protected]/0.5m/1m/1m 30°, surpassing prior methods, including those using semantic supervision.

Ablation studies reveal that:

• Using both position- and orientation-level negatives yields maximal recall.
• Removing the CLS token notably reduces recall at fine localization ([email protected] m).
• Positional ($1 \times 1$4 at [email protected] m) and angular ($1 \times 1$5 at R@1 m) perturbations are important for robust training.
• Best performance is observed with $1 \times 1$6 candidates, $1 \times 1$7, and 5 m $1 \times 1$8 5 m patches.

5. Analysis and Implications

The two-level contrastive scheme directly addresses the ambiguous pose hypotheses arising from architectural repetition:

• Position-level negatives drive the system to distinguish spatially nearby but geometrically similar regions, minimizing self-similarity confounds.
• Orientation-level negatives enforce heading sensitivity, critical in symmetric environments.
• Orientation-level contrast and angular perturbation offer slightly larger accuracy gains than positional mechanisms.

Importantly, the DINO V2 depth encoder transfers monocular depth estimation skill to the FLoc context without fine-tuning, and no semantic (e.g., room-type, door) annotations are required, achieving superior localization compared to state-of-the-art semantic-based approaches.

6. Limitations and Future Directions

DisCo-FLoc’s sequential ray regression and contrastive reranking architecture introduces some redundancy in feature extraction between the two stages. A plausible implication is that a unified, end-to-end framework could eliminate this inefficiency and further improve localization accuracy.

Potential avenues for extension include jointly integrating depth regression and contrastive disambiguation within a single module, streamlining computation and potentially exploiting cross-modality cues more effectively.

---

The DisCo-FLoc framework establishes a robust, annotation-free approach for monocular visual floorplan localization, coupling depth-aware geometry prediction with dual-level contrastive disambiguation. Its mathematical formulations, dual-stage design, and empirical performance on challenging benchmarks demonstrate significant improvements in both localization precision and orientation disambiguation (Meng et al., 5 Jan 2026).