New ML Pipeline Auto-Groups Vacation Rental Photos by Room

A new research paper presented at the KDD 2025 Two-sided Marketplace Optimization Workshop introduces a computationally efficient machine learning pipeline for organizing unstructured vacation rental image collections. The system tackles two key challenges: grouping images that belong to the same room (room scene discovery and grouping) and identifying bed types within bedroom groups. The pipeline integrates three components: a supervised room-type detection model, a supervised overlap detection model that measures similarity between image pairs, and a clustering algorithm that groups images based on similarity scores. Additionally, a Multimodal Large Language Model (MLLM) maps each bedroom group to the bed types specified in the property's metadata by analyzing visual content.

The researchers evaluated each model individually and the complete pipeline, finding it significantly outperforms established approaches like contrastive learning and clustering with pretrained embeddings. The system is designed for low latency and sample-efficient learning, making it suitable for real-time environments and platforms with limited labeled data. This work addresses a common pain point for travelers: understanding the spatial layout and sleeping configurations of vacation rentals from disorganized photo galleries. By automatically grouping identical rooms and identifying bed types, the pipeline could help rental platforms improve user experience and reduce booking uncertainty.