Room Scene Discovery and Grouping in Unstructured Vacation Rental Image Collections

TL;DR

This paper presents a machine learning pipeline for discovering and grouping room scenes in unstructured vacation rental images, enabling better spatial understanding and bed type identification, with high efficiency and accuracy.

Contribution

The authors introduce a novel, low-latency, sample-efficient machine learning pipeline for room scene discovery, grouping, and bed type identification in vacation rental images, outperforming existing methods.

Findings

Strong performance in room grouping accuracy

Effective bed type identification from images

Outperforms contrastive learning and pretrained clustering methods

Abstract

The rapid growth of vacation rental (VR) platforms has led to an increasing volume of property images, often uploaded without structured categorization. This lack of organization poses significant challenges for travelers attempting to understand the spatial layout of a property, particularly when multiple rooms of the same type are present. To address this issue, we introduce an effective approach for solving the room scene discovery and grouping problem, as well as identifying bed types within each bedroom group. This grouping is valuable for travelers to comprehend the spatial organization, layout, and the sleeping configuration of the property. We propose a computationally efficient machine learning pipeline characterized by low latency and the ability to perform effectively with sample-efficient learning, making it well-suited for real-time and data-scarce environments. The pipeline integrates a supervised room-type detection model, a supervised overlap detection model to identify the overlap similarity between two images, and a clustering algorithm to group the images of the same space together using the similarity scores. Additionally, the pipeline maps each bedroom group to the corresponding bed types specified in the property's metadata, based on the visual content present in the group's images using a Multi-modal Large Language Model (MLLM) model. We evaluate the aforementioned models individually and also assess the pipeline in its entirety, observing strong performance that significantly outperforms established approaches such as contrastive learning and clustering with pretrained embeddings.