Boundary Cross-Attention (BCA) Mechanism

• BCA is an attention mechanism that explicitly models boundary information to improve spatial reasoning and intra-class consistency in tasks like segmentation and floorplan generation.
• Its design leverages dedicated boundary representations and cross-attention formulations to balance hard geometric constraints with solution diversity.
• Empirical results demonstrate enhanced boundary precision and performance improvements across both generative models and semantic segmentation benchmarks.

Boundary Cross-Attention (BCA) is a specialized attention mechanism that enables neural networks to explicitly model, condition on, or integrate boundary information for improved spatial reasoning, geometric consistency, and intra-class feature consistency. BCA has been developed for both generative modeling tasks—such as floorplan generation with strong geometric constraints—and discriminative tasks like semantic segmentation, where accurate boundary localization is central to overall performance. Key contributions include the explicit aggregation of context along learned boundaries and the ability to enforce and balance hard boundary adherence versus solution diversity in generative tasks (Stoppani et al., 2 Feb 2026, Ma et al., 2021).

1. Architectural Principles and Core Mechanisms

The central idea underlying BCA is the separation and explicit modeling of boundary information as a primary conditioning signal for attention-based architectures. In the context of generative models, BCA introduces a dedicated branch for encoding structural boundaries (e.g., floorplan perimeter polygons) (Stoppani et al., 2 Feb 2026). For segmentation, BCA computes boundary-aware attention maps that aggregate context specifically along detected object boundaries, diverging from standard self-attention which aggregates globally (Ma et al., 2021).

Floorplan Generation (HouseDiffusion/DDPM BCA)

• Boundary Representation: Encode a sequence of $M$ polygon corner coordinates $\{b_1,\ldots,b_M\}$, each embedded via a linear projection ($W_\text{emb} \in \mathbb{R}^{d \times 2}$) with positional encoding.
• Boundary Self-Attention: Enrich corner representations by computing multi-head self-attention among the $M$ tokens, yielding context-aware embeddings $H^B$.
• Room-to-Boundary Cross-Attention: Take $N$ room-corner tokens $X = \{x_1,\ldots,x_N\}$, form cross-attention queries from these, and attend to the boundary representations, computing

$$\mathrm{BCA}(X,B) = \mathrm{softmax}\left( Q^X (K^B)^\top / \sqrt{d_k} \right) V^B.$$

• Integration: Output added to existing Transformer attention heads (Component-wise, Global, Relational), i.e.,

$$A_\text{out} = A_\text{CSA} + A_\text{GSA} + A_\text{RCA} + \mathrm{BCA}(X,B).$$

Semantic Segmentation (BCANet/Non-local BCA)

• Multi-Scale Boundary Extraction (MSB): Boundary maps are extracted at multiple backbone stages and fused into a $C_2$-channel feature map via convolutions and upsampling.
• Semantic and Boundary Features: Project backbone features ($\{b_1,\ldots,b_M\}$0) and fused MSB output ($\{b_1,\ldots,b_M\}$1) to a common embedding space using $\{b_1,\ldots,b_M\}$2 or $\{b_1,\ldots,b_M\}$3 convolutions.
• Cross-Attention Formulation: Compute similarity $\{b_1,\ldots,b_M\}$4 between boundary positions $\{b_1,\ldots,b_M\}$5 and semantic pixel $\{b_1,\ldots,b_M\}$6:

$\{b_1,\ldots,b_M\}$7

Aggregate features so every pixel receives a boundary-contextualized update:

$\{b_1,\ldots,b_M\}$8

with residual connections.

2. Mathematical Formulation of Boundary Cross-Attention

The core mathematical operation applies multi-head attention to selectively aggregate information using boundary features as keys/values and task-specific features (e.g., room-corner or semantic maps) as queries. In either setting, standard multi-head attention is given by

$\{b_1,\ldots,b_M\}$9

Generative Setting (Room $W_\text{emb} \in \mathbb{R}^{d \times 2}$0 Boundary):

• $W_\text{emb} \in \mathbb{R}^{d \times 2}$1
• $W_\text{emb} \in \mathbb{R}^{d \times 2}$2
• $W_\text{emb} \in \mathbb{R}^{d \times 2}$3

Segmentation (Boundary $W_\text{emb} \in \mathbb{R}^{d \times 2}$4 Semantic):

• Similarity and aggregation are computed as above, with boundary features as attention keys.

3. Training Objectives and Evaluation Metrics

Floorplan Generation

• Diffusion Loss: Standard L2 denoising as in DDPMs.
• Diversity Score (DS): Quantifies sample variability under fixed constraints, computed as trace of covariance of Inception-V3 features:

$W_\text{emb} \in \mathbb{R}^{d \times 2}$5

• Perceptual and Geometric Metrics:
  • FID (Fréchet Inception Distance): Realism vs. ground truth.
  • Boundary Compatibility (BC): Average Hausdorff distance between generated and ground-truth boundaries.
  • Graph Compatibility (GC): Adjacency adherence.

Segmentation

• Segmentation Loss: Cross-entropy on class logits.
• Boundary Loss: Binary cross-entropy on edge maps from MSB.
• Boundary-aware Segmentation Loss: Penalizes segmentation errors at detected boundaries.
• Auxiliary Loss: Intermediate supervision.
• Aggregate Loss: Weighted sum with $W_\text{emb} \in \mathbb{R}^{d \times 2}$6

4. Empirical Results and Ablation Studies

Generative Boundary Cross-Attention

• Boundary Compatibility (RPLAN, 512 samples):
  • Graph2Plan: BC = 0.11
  • Cons2Plan: BC = 0.06
  • HD+BCA: BC = 0.04 (best observed)
• Realism-Diversity Trade-off:
  • FID and BC improve with training, while DS collapses, indicating mode collapse.
  • At 10k steps: FID=25.98, BC=1.00, DS=52.69; at 400k: FID=10.74, BC=0.04, DS=33.59.
  • A plausible implication is that optimizing solely for realism and boundary fit reduces sample diversity.
• Out-of-Distribution Generalization:
  • Minor distributional shifts (Drift): Rapid adaptation, e.g., FID drops from 38.69 (0 shots) to 26.47 (20 shots).
  • Major geometry (Synthetic): FID remains elevated, e.g., 290.00 → 121.54.

Segmentation BCA (BCANet)

• Cityscapes (val, mIoU):
  • Dilated-FCN: 75.52
  • + MSB only: 74.95
  • + Single-scale BCA: 79.06
  • + Multi-scale BCA: 80.03
  • + MSB + BCA + boundary-aware loss: 80.92
• State-of-the-Art:
  • Cityscapes test: 81.7% mIoU (BCANet, ResNet-101)
  • ADE20K val: 45.62% mIoU, 82.35% pixel-accuracy
• Boundary/Interior F-score:

| Module | Boundary F | Interior F | |--------------------|------------|------------| | Dilated-FCN | 57.93 | 75.07 | | + SegFix | 61.70 | 75.26 | | + MSB-BCA | 60.38 | 77.02 | | + MSB-BCA+SegFix | 63.92 | 77.13 |

5. Applications and Key Insights

• Hard Constraint Enforcement: In generative models, BCA enables strict adherence to user-specified geometric boundaries (evidenced by lowest BC values), facilitating interactive design tools where custom lot geometries must be respected.
• Realism-Diversity Control: Classifier-free guidance with boundary dropout enables explicit tuning of adherence-diversity trade-offs through inference-time scaling parameter $W_\text{emb} \in \mathbb{R}^{d \times 2}$7; high $W_\text{emb} \in \mathbb{R}^{d \times 2}$8 enforces constraints, while lower values increase diversity (Stoppani et al., 2 Feb 2026).
• Segmentation Consistency: By aggregating context exclusively along detected boundaries, BCA increases intra-class consistency and reduces class confusion at object borders, outperforming other context modules such as Non-local, ASPP, and DNL (see mIoU comparisons) (Ma et al., 2021).
• Generalization Limitations: Out-of-distribution evaluation indicates that existing boundary-driven generative models remain reliant on training priors for room arrangement complexity. This suggests future advances would benefit from incorporating richer geometric reasoning mechanisms and countering mode collapse, as evidenced by the Diversity Score findings.

6. Comparison with Other Attention Mechanisms

BCA distinguishes itself by the explicit use of boundary features as attention keys, in contrast to ubiquitous self-attention schemes:

Module	Key Type	Cross/Self	Performance (Cityscapes mIoU)
ASPP	Global (dilated)	Self	79.29
Non-local	All tokens	Self	78.60
Criss-Cross (RCCA)	Criss-cross tokens	Self+Criss	78.44
DNL	Dense	Self	79.38
MSB-BCA (Ours)	Boundary features	Cross	80.03

This heightened focus on boundaries delivers superior overall accuracy and especially enhanced boundary precision (Ma et al., 2021).

7. Practical Implications and Future Directions

Boundary Cross-Attention modules have immediate utility in any pipeline that requires conditional geometric generation or precise discriminative boundary reasoning. Their modular formulation, compatibility with common backbones, and demonstrated effectiveness in both DDPM-based generative models and semantic segmentation architectures establish BCA as a primary paradigm for spatially-constrained neural modeling. However, challenges remain regarding sample diversity, mode collapse, and OOD robustness. A plausible implication is that BCA, when combined with adaptive guidance or alternative boundary encodings, will form the basis for future systems balancing fidelity and creativity in structural design and dense labeling tasks (Stoppani et al., 2 Feb 2026, Ma et al., 2021).