1# RECORD 5bcd886348a12f52aa055354a9647ba5c4fa64c42# TITLE: Giving Eyes to Automated Valuation Models: Computer Vision Assessment of Interior Home Condition3# YEAR: 2026.0 VENUE: Journal of real estate finance and economics DOI: 10.1007/s11146-026-10071-w4# AUTHORS: V. Almaas; A. Oust; A. J. Pollestad5# PAGES: 35 CANDIDATE TABLE PAGES (1-based): [19, 21, 22, 31]6# ABSTRACT: 7 Housing condition is a key determinant of property value, yet it is often omitted from modern automated valuation models (AVMs). This study develops a computer vision framework to extract room-specific condition signals from listing images and examines their economic relevance for housing valuation. Using 15,702 condominium transactions from Oslo, Norway (2022–2023), matched with 393,914 images, we classify images by room type, estimate condition on a 1–5 scale, and incorporate these measures into hedonic linear regression and XGBoost AVMs. The results show that computer-vision-based condition improves predictive performance, with gains comparable to those from human-labeled condition. Moreover, housing condition is positively associated with sales prices, with stronger effects for kitchens and bathrooms and in lower-priced homes. Explainable AI analysis further indicates that the models rely on intuitive and economically meaningful visual patterns.8910===== PAGE 1 (score 7, numeric-paren 0, table-words 0, var-words 7) =====1112The Journal of Real Estate Finance and Economics13https://doi.org/10.1007/s11146-026-10071-w14 15 Arne Johan Pollestad16arne.j.pollestad@ntnu.no17Vebjørn Almaas18vebjorn.almaas@ntnu.com19Are Oust20are.oust@ntnu.no211 22NTNU Business School, Norwegian University of Science and Technology, Trondheim, 23Norway242 25Norwegian School of Economics, Bergen, Norway26Received: 23 September 2025 / Accepted: 26 June 202627© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 28202629Giving Eyes to Automated Valuation Models: Computer 30Vision Assessment of Interior Home Condition31Vebjørn Almaas1,232 · Are Oust133 · Arne Johan Pollestad134Abstract35Housing condition is a key determinant of property value, yet it is often omitted 36from modern automated valuation models (AVMs). This study develops a computer 37vision framework to extract room-specific condition signals from listing images and 38examines their economic relevance for housing valuation. Using 15,702 condomin39ium transactions from Oslo, Norway (2022–2023), matched with 393,914 images, 40we classify images by room type, estimate condition on a 1–5 scale, and incorpo41rate these measures into hedonic linear regression and XGBoost AVMs. The results 42show that computer-vision-based condition improves predictive performance, with 43gains comparable to those from human-labeled condition. Moreover, housing condi44tion is positively associated with sales prices, with stronger effects for kitchens and 45bathrooms and in lower-priced homes. Explainable AI analysis further indicates that 46the models rely on intuitive and economically meaningful visual patterns.47Keywords Automated valuation model (AVM) · Image segmentation · Condition · 48Computer vision · Explainable AI · Real estate valuation49JEL-Classification C55 · R21 · R31505152===== PAGE 19 (score 11, numeric-paren 0, table-words 3, var-words 3) =====53541 355Giving Eyes to Automated Valuation Models: Computer Vision…56Table 4 AVM performances with condition variables57AVM58Condition variable59MAPE60MdAPE61R262> 10%63HLR64Excluding condition6512.75%6610.11%670.6726849.70%69HLR70Computer vision condition7112.42%729.91%730.6877450.55%75HLR76Human condition7712.41%789.75%790.6908051.19%81XGBoost82Excluding condition838.03%845.76%850.8478673.43%87XGBoost88Computer vision condition897.71%905.49%910.8559275.68%93XGBoost94Human condition957.49%965.34%970.8639875.76%99This table shows the predictive performances of the two AVMs, a hedonic linear regression (HLR) and 100an XGBoost, under three different specifications. First, we train our model only on traditional housing 101data. Secondly, we include thehi computer vision condition variable to the input data. Finally, we replace 102the computer vision condition variable with human condition labels from the crowdsourcing process. 103The best-performing model(s) for each AVM is highlighted in bold. Source: Author calculations using 104Boligmappa AS housing data, January 2022–December 2023105Fig. 7 Average condition levels across city districts. Notes: This figure displays the spatial distribution 106of average room condition across the administrative districts of Oslo. A blue gradient with a corre107sponding numerical scale is adopted to represent the average condition levels, where lighter (darker) 108shades indicate worse (better) condition. The 15 district names are labeled on the map. Oslo Sentrum 109(CBD) is marked in gray due to the absence of data in our sample. Source: Author calculations using 110Boligmappa AS housing data, January 2022–December 2023111 112113114===== PAGE 20 (score 11, numeric-paren 0, table-words 2, var-words 7) =====1151161 3117V. Almaas et al.118economically meaningful information that is not fully captured by standard housing 119characteristics.120For the HLR model, MAPE declines from 12.8% to 12.4% when computer-121vision-based condition is included, while R2 increases from 67.2% to 68.7%. These 122improvements suggest that the condition measures capture variation in housing qual123ity that is priced in the market but omitted from conventional hedonic specifications. 124Importantly, the performance gains are similar when using human-labeled condition, 125indicating that the computer vision measures recover much of the same information.126A similar pattern is visible for the XGBoost model. Including computer-vision-127based condition reduces MAPE from 8.0% to 7.7%, while the addition of human con128dition labels yields a further improvement (MAPE of 7.5%). Across the two models, 129this suggests that the information extracted by the computer vision models resembles 130that provided by human assessments of interior condition.131A reduction in average valuation error of 0.4% points equals a reduction of around 132NOK 24,000 for the average home with a price per m2 of NOK 90,000 and a size of 13366 m2. By extracting valuation-relevant information from listing images, computer 134vision models may therefore help reduce omitted variable bias in pricing models and 135improve the accuracy of housing valuations at an economically meaningful scale.136Hedonic Pricing of Condition Signals137Furthermore, we can examine the estimated coefficients of the condition variables in 138the HLR model, reported in Table 5. Columns 1–2 include city district fixed effects, 139while columns 3–4 employ more granular postal code fixed effects. Moving to finer 140spatial fixed effects allows us to absorb increasingly detailed location-specific price 141variation, thereby reducing concerns that the estimated condition effects are driven 142by unobserved neighborhood characteristics. The coefficients are thus identified from 143within-location variation in condition across otherwise comparable properties, pro144viding a cleaner estimate of how computer-vision-based signals of interior condition 145are associated with predicted sales prices.146With few exceptions, primarily insignificant estimates for low-condition bed147rooms, the results consistently indicate that higher condition levels are associated 148with higher property prices. Economically, the effects are substantial, as a one-level 149increase in average condition is associated with a price premium of approximately 1508.2% to 8.7%. This suggests that visual interior condition signals capture a meaning151ful dimension of housing quality that adds value to price models. The stability of the 152condition coefficients across the different specifications also suggests that the results 153are not driven by spatial sorting of higher-condition homes into more expensive areas.154To illustrate the magnitude of the effect, consider a representative condominium in 155our sample, with an average size of 66 m² and a price of approximately NOK 90,000 156per square meter, corresponding to a total value of roughly NOK 5.9 million. An 8% 157increase in price therefore implies an increase of almost NOK 500,000. These esti158mates highlight that relatively small improvements in measured condition translate 159into economically significant differences in predicted values.160The room-specific condition variables further allow us to decompose the average 161effect across different parts of the dwelling. For bathrooms, condition is statistically 162163164===== PAGE 21 (score 46, numeric-paren 27, table-words 5, var-words 6) =====1651661 3167Giving Eyes to Automated Valuation Models: Computer Vision…168distinguishable from the reference category (level 4) across all levels, with clear val169uation penalties for lower condition and premiums for higher ones. A similar pattern 170emerges for kitchens and living rooms, with the effects being particularly pronounced 171for kitchens. This suggests that kitchens may be the room type for which the image-172based measures contain the most valuation-relevant information. Economically, the 173magnitudes are substantial as a level 1 kitchen is associated with a valuation discount 174of approximately 17.2% to 18.1%, while a level 5 kitchen commands a premium of 1757.8% to 7.9% relative to the reference level. The results generally imply that visual 176condition measures capture meaningful heterogeneity in interior condition across 177room types.178Next, we conduct a series of heterogeneity analyses to examine whether the results 179reported in Table 5 vary across subsamples of condominiums. First, we split the sam180ple into quartiles based on transaction prices, comparing the lowest and highest price 181segments. This allows us to assess whether the valuation impact of image-based con182Table 5 Hedonic linear regression condition effects183Condition variable184(1)185(2)186(3)187(4)188Room-specific189Average190Room-specific191Average192Dependent variable193in line-eq-IEq12 194in line-eq-IEq13195in line-eq-IEq14196in line-eq-IEq15197CondAverage1980.087***1990.082***200CondBath = 1201-0.062**202-0.057**203CondBath = 2204-0.050***205-0.041***206CondBath = 3207-0.011***208-0.014***209CondBath = 52100.027***2110.025***212CondBed = 1213-0.099***214-0.087***215CondBed = 2216-0.043***217-0.031***218CondBed = 3219-0.004220-0.003221CondBed = 52220.1362230.145*224CondLiv = 1225---226---227CondLiv = 2228-0.045**229-0.039**230CondLiv = 3231-0.011***232-0.010***233CondLiv = 52340.021235-0.015236CondKit = 1237-0.172***238-0.181***239CondKit = 2240-0.087***241-0.091***242CondKit = 3243-0.036***244-0.037***245CondKit = 52460.079***2470.078***248Dwelling Attributes249Yes250Yes251Yes252Yes253Year x Quarter FE254Yes255Yes256Yes257Yes258District FE259Yes260Yes261No262No263Postal code FE264No265No266Yes267Yes268Adjusted R22690.69812700.69522710.7722720.771273Observations27415,70227512,08427615,70227712,084278The coefficients represent the percentage effect on price per m2 from a hedonic linear regression model 279for different condition variables. CondAverage refers to an aggregated condition variable across room 280types, while CondBath, CondBed, CondLiv, and CondKit represent the room-specific conditions. All 281conditions are measured on a scale from 1–5. The models have fixed effects for the interaction between 282year and quarter, as well as city district fixed effects in models 1–2 and postal code fixed effects in 283models 3–4. *p<0.1; **p<0.05; ***p<0.01. Source: Author calculations using Boligmappa AS housing 284data, January 2022–December 2023285286287===== PAGE 22 (score 68, numeric-paren 41, table-words 7, var-words 5) =====2882891 3290V. Almaas et al.291dition signals differs across the price distribution, for instance if interior quality is 292more strongly capitalized in high-end markets. Second, we split the sample based on 293the number of listing images, comparing the quartile with the fewest images to the 294quartile with the most. The number of images serves as a proxy for marketing effort 295and information availability, enabling us to test whether the valuation effects of con296dition signals depend on the richness of visual information provided.297Finally, we augment the baseline specification with broker fixed effects. Since 298listing images are produced by professional real estate brokers, this specification 299controls for systematic differences in photography quality, staging, and presenta300tion across professionals. It thereby mitigates concerns that the estimated condition 301effects may partly reflect broker-specific practices correlated with both image quality 302and property values.303The results are shown in Table 6. The table reports our preferred specification 304with postal code fixed effects, while the corresponding district fixed effects estimates 305are reported in Appendix Table 7. In the full-sample specification with broker fixed 306Table 6 Heterogeneity analysis of condition effects307Sample308(1)309(2)310(3)311(4)312(5)313Cheap homes314Expensive homes315Few pictures316Many pictures317Full318Dependent variable319in line-eq-IEq16320in line-eq-IEq17321in line-eq-IEq18322in line-eq-IEq19323in line-eq-IEq20324CondBath = 1325-0.048*3260.051327-0.054*328-0.188***329-0.068***330CondBath = 2331-0.030***332-0.01333-0.034***334-0.047***335-0.042***336CondBath = 3337-0.020***3380.011**339-0.010*340-0.021***341-0.014***342CondBath = 53430.024**3440.0173450.030**3460.0073470.023***348CondBed = 1349-0.053**350---351-0.016352-0.157353-0.083***354CondBed = 2355-0.031***3560.0123570.011358-0.047359-0.038***360CondBed = 3361-0.017***3620.0003630.010*3640.003365-0.002366CondBed = 53670.0093680.0673690.1263700.1103710.146*372CondLiv = 1373---374---375---376---377---378CondLiv = 2379-0.022380-0.0153810.0013820.001383-0.036**384CondLiv = 3385-0.015***386-0.007387-0.009388-0.009389-0.010***390CondLiv = 5391-0.0573920.0203930.0343940.017395-0.013396CondKit = 1397-0.079***398-0.170399-0.184***400-0.186401-0.148***402CondKit = 2403-0.059***404-0.025405-0.089***406-0.055*407-0.077***408CondKit = 3409-0.021***410-0.009411-0.040***412-0.033***413-0.031***414CondKit = 5415-0.0064160.042***4170.083**4180.061***4190.076***420Dwelling Attributes421Yes422Yes423Yes424Yes425Yes426Year x Quarter FE427Yes428Yes429Yes430Yes431Yes432Postal code FE433Yes434Yes435Yes436Yes437Yes438Broker FE439No440No441No442No443Yes444Adjusted R24450.7144460.2834470.8614480.6944490.756450Observations4513,9264523,9264533,1104543,10545515,702456The coefficients represent the percentage effect on price per m2 from a hedonic linear regression model 457for different condition variables. CondBath, CondBed, CondLiv, and CondKit represent the room-458specific conditions. All conditions are measured on a scale from 1–5. The models have fixed effects 459for the interaction between year and quarter, as well as postal code fixed effects. The model in column 4605 also includes broker fixed effects. *p<0.1; **p<0.05; ***p<0.01. Source: Author calculations using 461Boligmappa AS housing data, January 2022–December 2023462463464===== PAGE 23 (score 13, numeric-paren 0, table-words 1, var-words 11) =====4654661 3467Giving Eyes to Automated Valuation Models: Computer Vision…468effects in Column 5, level 1 kitchens are associated with a valuation discount of 46914.8%, while level 2 and level 3 kitchens are discounted by 7.7% and 3.1%, respec470tively. Bathrooms display a similar ordering, with lower condition levels associated 471with discounts and level 5 bathrooms associated with a positive premium. These 472patterns suggest that the image-based variables capture valuation-relevant signals of 473interior condition rather than merely reflecting broker-specific marketing practices.474The heterogeneity analysis in Columns 1–4 further indicates that condition effects 475are more pronounced among lower-priced homes. In the cheapest quartile, poor 476kitchen, bathroom, and bedroom condition is consistently associated with lower 477prices, whereas the corresponding penalties are weaker and less precisely estimated 478among the most expensive homes. This suggests that visible interior deficiencies may 479be more informative for valuation in lower-priced market segments, where required 480renovation costs represent a larger share of the total property value.481The results by number of listing images are more mixed. While some estimates 482suggest stronger penalties for poor condition among homes with richer visual docu483mentation, these patterns are less robust once postal code fixed effects are included. 484We therefore interpret this evidence cautiously.485The HLR coefficients from Table 5 align prior with AVM research that emphasize 486the importance of incorporating detailed, room-specific condition measures to explain 487sales prices (Koch et al., 2024; Miller et al., 2018). The strong price premiums for 488high-standard bathrooms and the penalties for substandard conditions supports the 489argument made by Oust et al. (2023) that interior quality significantly affects property 490valuation. Furthermore, the differential impact across price segments, where poor 491bathroom condition appears to reduce sale prices more for low-end properties, reso492nates with the broader evidence from Mamre and Sommervoll (2024) and Francke 493and van de Minne (2017), who show that both renovations and lack of maintenance 494have substantial effects on sales prices.495Non-Linear Model Interpretation and Explainability496After assessing the hedonic pricing of condition signals, explainable AI methodology 497allows for similar analysis for the non-linear XGBoost AVM. These are displayed in 498Fig. 8. Both Panel A, which displays the four room-specific condition variables, and 499Panel B, which shows the average condition level, reveal a clear positive relationship 500between condition and property price. This is evident since higher condition values 501(in red colors) have positive marginal effects on predicted sales price (positive SHAP 502values).503To further examine the marginal effect of condition, we present SHAP dependence 504plots in Panels A–E of Fig. 9. In these plots, the horizontal axis indicates condition 505level, the vertical axis shows the marginal price impact (SHAP value) of condition, 506and coloring reflects the actual sales price. Ignoring coloration for the moment, the 507relationship between condition and marginal price impact is consistently positive 508across all room types: higher condition levels lead to higher predicted prices from the 509XGBoost AVM, holding other variables constant.510As indicated by the confusion matrices in Figs. 3, 4, 5, 6, and 15 in the Appendix, 511a large share of homes was classified at levels 3 and 4. This distribution explains why 512513514===== PAGE 31 (score 61, numeric-paren 39, table-words 6, var-words 2) =====5155161 3517Giving Eyes to Automated Valuation Models: Computer Vision…518However, an important difference is that the condition coefficients are more fre519quently statistically significant in the subsamples based on the number of images. In 520particular, the estimates for homes with few and many pictures tend to be larger in 521magnitude and more precisely estimated under the district-level specification. This 522likely reflects the broader spatial controls, which leave more variation in prices to be 523explained by observable characteristics such as condition. The results suggest that 524while the main findings are robust to the choice of spatial fixed effects, the magni525tude and statistical significance of the condition effects are somewhat sensitive to 526the level of spatial aggregation. The more conservative postal code specification in 527Table 6 absorbs finer location-specific price variation, reducing the risk that condi528tion coefficients capture unobserved neighborhood characteristics. By contrast, the 529district-level specification in Table 7 may partially assign residual spatial variation to 530condition, leading to stronger and more precisely estimated effects.531Importantly, the consistency in the direction and economic interpretation of the coef532ficients across both specifications supports the conclusion that the computer-vision-533based condition measures capture valuation-relevant information on interior quality, 534rather than merely reflecting spatial sorting or differences in marketing practices.535Table 7 Heterogeneity analysis of condition effects536Sample537(1)538(2)539(3)540(4)541(5)542Cheap homes543Expensive homes544Few pictures545Many pictures546Full547Dependent variable548in line-eq-IEq21 549in line-eq-IEq22 550in line-eq-IEq23 551in line-eq-IEq24 552in line-eq-IEq25553CondBath = 1554-0.0345550.010556-0.050557-0.218***558-0.062**559CondBath = 2560-0.039***561-0.019562-0.043***563-0.051***564-0.049***565CondBath = 3566-0.0075670.01**568-0.010*569-0.022***570-0.012***571CondBath = 55720.0135730.021*5740.036**5750.0045760.026***577CondBed = 1578-0.0504*579---580-0.016581-0.222582-0.095***583CondBed = 25840.035***5850.0225860.010587-0.070*588-0.041***589CondBed = 3590-0.023***5910.0025920.009593-0.003594-0.003595CondBed = 55960.0355970.0425980.1195990.1326000.146*601CondLiv = 1602---603---604---605---606---607CondLiv = 2608-0.0306090.0086100.0136110.010612-0.044**613CondLiv = 3614-0.013***615-0.008616-0.012**617-0.004618-0.010***619CondLiv = 56200.0136210.0056220.0336230.0516240.018625CondKit = 1626-0.083***627-0.157628-0.191***629-0.419630-0.168***631CondKit = 2632-0.059***633-0.024634-0.085***635-0.077**636-0.083***637CondKit = 3638-0.022***639-0.012**640-0.037***641-0.034***642-0.034***643CondKit = 5644-0.0296450.052***6460.0896470.066***6480.081***649Dwelling Attributes650Yes651Yes652Yes653Yes654Yes655Year x Quarter FE656Yes657Yes658Yes659Yes660Yes661District FE662Yes663Yes664Yes665Yes666Yes667Broker FE668No669No670No671No672Yes673Adjusted R26740.6086750.1416760.8256770.5926780.707679Observations6803,9266813,9266823,1106833,10568415,702685The coefficients represent the percentage effect on price per m2 from a hedonic linear regression model 686for different condition variables. CondBath, CondBed, CondLiv, and CondKit represent the room-specific 687conditions. All conditions are measured on a scale from 1–5. The models have fixed effects for the interaction 688between year and quarter, as well as district fixed effects. The model in column 5 also includes broker fixed 689effects. *p<0.1; **p<0.05; ***p<0.01. Source: Author calculations using Boligmappa AS housing data, January 6902022–December 2023691692693===== PAGE 32 (score 3, numeric-paren 0, table-words 0, var-words 3) =====6946951 3696V. Almaas et al.697Distribution of Classified Condition Levels698Figure 15 summarizes the condition variables derived from the CNN computer vision 699classifications. Across all four room types, the majority of images are classified at 700condition levels 3 and 4, indicating generally high maintenance standards. Overall, 701condition levels 1, 2, and 5 are underrepresented, suggesting that most Oslo condo702miniums fall within the mid- to high-quality range.703Acknowledgements We thank the participants of the American Real Estate Society (ARES) Spring Con704ference 2026 for their valuable comments and discussions. We are also grateful to the two anonymous 705reviewers for their thoughtful and constructive feedback, which substantially improved the quality and 706clarity of the manuscript. An earlier version of this work was submitted as a master’s thesis by V.A. at the 707Norwegian School of Economics in the fall of 2024. The present article has undergone extensive revisions 708and extensions and should be regarded as an independent research contribution. Any remaining errors or 709omissions are solely the responsibility of the authors.710Author Contributions All authors contributed to the conceptualization and methodology. V.A. performed 711the data cleaning and data analysis. A.O. and A.P. wrote the main manuscript text. All authors participated 712in revising the manuscript and approved the final version.713Fig. 15 Classification frequency. Notes: This figure shows the distributions of condition classification 714frequencies for four different room types. All condition assessments are classified using computer vi715sion technology. Condition is assessed on a scale from 1–5, where higher numbers represent better con716dition. Source: Author calculations using Boligmappa AS housing data, January 2022–December 2023717 718