Source-linked AI summary
Vision-based Real Estate Price Estimation
Omid Poursaeed, Tomas Matera, Serge Belongie
TL;DR
Existing Zillow and Redfin automated valuation methods do not consider interior and exterior appearance, although the paper identifies appearance as relevant to house value. The authors estimate luxury levels from interior and exterior photos and combine the visual information with house metadata to estimate value. The network achieves 5.8 percent, better than the Zestimate’s 7.9 percent median error rate.
Problem
Existing Zillow and Redfin automated valuation methods do not consider interior and exterior appearance, although the paper identifies appearance as relevant to house value.
Source: §1 · p. 1Method
The authors estimate luxury levels from interior and exterior photos and combine the visual information with house metadata to estimate value.
Source: §3 · p. 3, §3.4 · p. 7Results
The network achieves 5.8 percent, better than the Zestimate’s 7.9 percent median error rate.
Source: §4 · p. 8Takeaways & Limitations
The reported result supports using home images alongside other available information in automated valuation.
Source: §4 · p. 8, §5 · p. 8The room-category classifier sometimes misclassifies empty rooms and rooms combining elements from different categories.
Source: §3.1 · p. 4
Abstract
from arXiv · showhide
Since the advent of online real estate database companies like Zillow, Trulia and Redfin, the problem of automatic estimation of market values for houses has received considerable attention. Several real estate websites provide such estimates using a proprietary formula. Although these estimates are often close to the actual sale prices, in some cases they are highly inaccurate. One of the key factors that affects the value of a house is its interior and exterior appearance, which is not considered in calculating automatic value estimates. In this paper, we evaluate the impact of visual characteristics of a house on its market value. Using deep convolutional neural networks on a large dataset of photos of home interiors and exteriors, we develop a method for estimating the luxury level of real estate photos. We also develop a novel framework for automated value assessment using the above photos in addition to home characteristics including size, offered price and number of bedrooms. Finally, by applying our proposed method for price estimation to a new dataset of real estate photos and metadata, we show that it outperforms Zillow's estimates.
1. Introduction
Automated valuation methods use home characteristics and public data but omit interior and exterior appearance, which the paper identifies as relevant to market value. The authors propose using real-estate imagery to estimate luxury and improve valuation, and introduce datasets and a Zillow comparison.
1. Introduction
- The paper introduces a method for incorporating interior and exterior appearance into real-estate price estimation, which existing Zillow and Redfin methods do not consider.It uses deep neural networks and crowdsourced information to derive luxury-related information from imagery.Source: §1 · p. 1
- The authors release photos and metadata for 9,000 Zillow houses and report that applying their valuation method to this dataset outperforms Zillow’s estimates.Source: §1 · p. 2
- The authors release a dataset of photos and metadata for 9,000 Zillow houses and report that their valuation method outperforms Zillow’s estimates.Source: §1 · p. 2
- The paper also releases 140,000 Houzz interior-design photos, categorized by room type, and presents a visualization clustering photos with similar luxury levels.Source: §1 · p. 2
2. Related Work
The review spans metadata-based valuation and computer-vision research, identifying visual appearance as an aspect omitted from existing valuation methods. It also surveys scene datasets and work on visual style.
2.1. Automated Valuation Methods
- The comparables model adjusts nearby sale prices; regression approaches relate home characteristics to value, while reviewed valuation methods omit visual features.The review also covers linear and hedonic pricing, neural networks, and fuzzy logic.Source: §2.1 · p. 2
2.2. Convolutional Neural Networks (ConvNets)
- ConvNets are presented as achieving state-of-the-art performance across image recognition, segmentation, object detection, and generative modeling.Source: §2.2 · p. 2
2.3. Scene Understanding
- Scene-understanding resources include Scene15, Places, and LSUN [30]; the review maps scene classes to real-estate room categories and compares category image counts.Source: §2.3 · pp. 2–3, Table 1 · p. 3
2.4. Visual Design
- Visual style is difficult to define rigorously; prior work covers style prediction [12], visual search, image similarity, furniture attributes, and scene representation.Source: §2.4 · p. 3
3. Our Approach
The approach treats luxury as a visual signal related to a home’s market value and combines photo-derived luxury information with house metadata. It uses crowdsourced labels and image classifiers to build house representations for price estimation.
3.1. Classifying Photos Based on Room Categories
- A DenseNet trained on more than 200k images classifies seven room categories at 91% test accuracy, enabling comparisons within room types.The dataset combines Places, Houzz, and Google image-search photos.Source: §3.1 · p. 3
3.2. Luxury Level Estimation
- Crowdsourced luxury comparisons produce an embedding whose clusters guide the number of luxury classes and the choice of representative images.Images with similar luxury levels cluster near one another in the embedding.Source: §3.2 · p. 4, §3.3 · p. 5
3.3. Crowdsourcing Framework
- Workers compare each probe with nine uniformly sampled photos, then classify probes against anchor images representing eight luxury levels.Tutorial rounds and catch trials were used to support attentive worker performance.Source: §3.3 · pp. 4, 7, Figure 4 · p. 6
3.4. Price Estimation Network
- A kernel support vector regression model maps each house’s concatenated room-luxury and normalized metadata vector to its purchase-price estimate.Recently sold houses provide the ground-truth prices used to train the model.Source: §3.4 · p. 7
4. Results and Discussion
On a test set of recently sold Zillow houses, the proposed method had lower median error than the Zestimate. Ablations report the effects of removing visual information, room classification, or luxury-based training, and of fine-tuning end-to-end.
4. Results and Discussion
- 5.8% median error on the 1,000-house test set beat Zestimate’s 7.9% median error.The authors attribute the improvement to augmenting the input data with images.Source: §4 · pp. 7–8
4.1. Ablation Studies
- 8.0% median error using metadata alone was close to the 7.9% Zestimate error rate.The authors note that Zestimate is included in the metadata and that the regression network nearly ignores other metadata elements.Source: §4.1 · p. 8
- 6.7% median error without the room classifier showed the importance of room-specific comparisons for fine-grained luxury signals.Source: §4.1 · p. 8
- Direct price regression yielded 6.6% median error, or 6.4% after fine-tuning the feature extractor.The authors attribute the lower performance to limited housing-price data and features that correlate poorly with the value images add.Source: §4.1 · p. 8
- End-to-end fine-tuning improved median error by 0.2% after the model parts were trained separately.The authors say this allows better information flow from regression to luxury estimation.Source: §4.1 · p. 8
5. Conclusion
The paper presents an algorithm that incorporates residential-property appearance into price estimation, using crowdsourced photo labels and house metadata. It reports better estimates than Zillow’s and identifies several directions for future research.
5. Conclusion
- The method estimates luxury from interior and exterior photos, then relates photos and metadata to the purchase prices of recently sold houses.The authors use crowdsourcing to extract luxury-related information and train convolutional neural networks for photo classification and price estimation.Source: §5 · p. 8
- The authors report that their algorithm provides better value estimates than Zillow’s estimates.The conclusion does not provide the comparison’s numerical result.Source: §5 · p. 8, Table 2 · p. 8
- Future research directions include assessing staging, comparing furniture styles, and developing an interface to help users select property photos.Other proposed directions include evaluating design photos aesthetically.Source: §5 · p. 8