Xelidoni: TREMA

Motivation

‍In the home rental industry, current damage assessment practices rely heavily on human inspectors, involving manual documentation and expert judgment. This process is time-consuming—often taking several hours per property—and incurs significant costs, especially when conducted at scale. To address these inefficiencies, our goal is to develop an AI-based damage assessment system that leverages visual recognition and anomaly detection techniques to automate inspections, reduce appraisal times from hours to minutes, and lower costs for property managers without compromising accuracy.

Proposed Approach / Solution

‍We explored the feasibility of leveraging AI to automatically detect damages incurred during home rentals by performing the following tasks:

1. Review of state-of-the-art damage detection methods and relevant public datasets: We explored three main damage detection approaches, including supervised defect classification, unsupervised anomaly detection, and language-guided anomaly detection.
2. Analysis of Xelidoni’s dataset and preliminary experiments using AnomalyCLIP: We propose building a dedicated dataset comprising approximately 5,000 normal and 500–750 anomalous images, covering a balanced range of high-, medium-, and low-priority indoor damages, annotated with anomaly labels and some bounding boxes. To ensure robustness and compatibility with public benchmarks, the dataset will include diverse scenes under different lighting conditions, varying damage severities, multiple viewing angles, and high-resolution images.

In addition, we tested AnomalyCLIP on normal and defective images of furnished flats, achieving strong zero-shot detection but encountering minor false positives on reflective surfaces. Since the model was trained on industrial datasets, fine-tuning it on a dedicated home damage dataset could improve performance. Future work could explore pretraining strategies and refining CLIP-based methods for better differentiation of defect types.

‍

‍

Impact

‍The AI-based damage assessment has the potential to transform the home rental industry by automating a traditionally time-consuming and costly inspection process. By leveraging visual recognition and anomaly detection techniques, the project aims to reduce inspection times from hours to minutes while lowering associated costs for property managers. This approach not only improves operational efficiency but also enhances the consistency and accuracy of damage appraisals across rental properties.