This repository contains two models trained for the research paper: "Unlocking Floorplan Retrieval with Design Intent via Contrastive Multimodal Learning".
These models are designed to retrieve architectural floorplans from a database based on a source image and a natural language instruction describing a desired change. This enables a more intuitive and goal-driven search for architects and designers.
Model Details
Two architectures were trained for this task using a triplet contrastive learning framework. The goal is to learn a shared embedding space where a query (source image + text instruction) is closer to a positive target image (that satisfies the instruction) than to a negative image.
1. CLIP-MLP-Floorplan-Retriever (Recommended)
This model uses the pre-trained multimodal embeddings from CLIP (ViT-B/32). The image and text embeddings are concatenated and passed through a simple MLP for fusion. This model demonstrated superior performance in both quantitative metrics and user studies.
Training Loss: TripletMarginWithDistanceLoss with Cosine Similarity (margin=0.2)
2. BERT-ResNet-CA-Floorplan-Retriever
This model uses separate pre-trained encoders for image and text. A cross-attention module is used to fuse the features, allowing the image representation to attend to linguistic cues from the instruction.
Image Encoder: ResNet50
Text Encoder: BERT (base-uncased)
Fusion: Cross-Attention Module
Training Loss: TripletMarginLoss with L2 Euclidean Distance (margin=1.0)
How to Use
You can use these models to get a fused embedding for a (floorplan, instruction) pair. You can then compare this embedding (e.g., using cosine similarity) against a pre-computed database of floorplan embeddings to find the best match.
