Disparity Estimates Research Articles

Observing Context Improves Disparity Estimation when Race is Unobserved

In many domains, it is difficult to obtain the race data that is required to estimate racial disparity. To address this problem, practitioners have adopted the use of proxy methods which predict race using non-protected covariates. However, these proxies often yield biased estimates, especially for minority groups, limiting their real-world utility. In this paper, we introduce two new contextual proxy models that advance existing methods by incorporating contextual features in order to improve race estimates. We show that these algorithms demonstrate significant performance improvements in estimating disparities, on real-world home loan and voter data. We establish that achieving unbiased disparity estimates with contextual proxies relies on mean-consistency, a calibration-like condition.

Housing-related disparities in receipt of breast cancer screening among Medicaid beneficiaries.

76 Background: People experiencing or at risk of homelessness (PEH) are often diagnosed with advanced breast cancer, suggesting barriers to preventive care and screening. However, there is limited data that examine screening among PEH, undermining efforts to characterize and address potential disparities in screening. To address this gap, we investigated housing-related disparities in receipt of breast cancer screening in Medicaid, which disproportionately serves low-income PEH. Methods: We used a novel linkage of Pennsylvania Medicaid data with encounter-level data on housing-related services (Housing Management Information System; HMIS) from 2011 to 2017. We compared rates of biennial mammography between a cohort of Medicaid women receiving housing-related services in HMIS (i.e., a proxy for experience/risk of homelessness) and randomly sampled comparison beneficiaries without any housing services use. We estimated both unadjusted and adjusted disparities in receipt of biennial mammography (controlled for county and year fixed effects, demographic and Medicaid eligibility characteristics, and baseline healthcare utilization and comorbidities). We also examined heterogeneity in disparities by beneficiaries in the following HMIS subgroups: 1) receiving supportive housing, 2) experiencing chronic homelessness (i.e., 12+ months of homelessness), and 3) unstably housed (housed but receiving homelessness prevention services). Results: We identified 1,792 Medicaid beneficiaries with an HMIS encounter and 71,664 comparison beneficiaries. Those in the HMIS cohort were more likely to be younger, non-Hispanic Black, eligible for Medicaid due to disability, and have a higher number of inpatient visits and comorbidities. The unadjusted proportion receiving mammogram was 44.8% and 50.1% in the HMIS vs. comparison cohort, with an overall disparity of 5.3 percentage-points (pp, P <.001). This disparity estimate remained was robust to adjusting for covariates (-4.3 pp; P < .001). We found the largest adjusted disparity when comparing beneficiaries experiencing chronic homelessness (-8.5 pp; P = .038) relative to the comparison cohort. In contrast, the disparity was smaller among women receiving supportive housing (-3.7 pp; P = .098) and unstably housed women (-3.3 pp; P = .071). Conclusions: There was a significant disparity in the utilization of mammography among women PEH, especially those who experience chronic homelessness, in Pennsylvania Medicaid. Our finding illuminates the health consequences of housing insecurity and underscore a need for programs providing adequate housing, such as supportive housing, to address disparities in screening. In the meantime, greater efforts should be dedicated to improving the uptake of screening among at-risk populations at the point of contact with the healthcare system (e.g. when PEH engage with care).

Classifying Cockroaches According to Forewings: Pitfalls and Implications for Fossil Systematics

A reliable character system is crucial to taxonomy and systematics, and it promises valid downstream inferences, e.g., estimates of diversity and disparity, reconstruction of evolutionary history, and even stratigraphic correlations. Modern taxonomy and systematics of extant cockroaches requires an integrative study involving multiple lines of evidence with emphasis on genital and reproductive characteristics and molecular data. In contrast, many fossil cockroach taxa published recently are based solely on forewings. Many studies have shown that forewing-based taxa are questionable. In order to find out how much of the phylogenetic signal we could ascertain from venational similarity, and how confident we could be, this study used forewing characters to reconstruct phylogenies of the genera of well-recognized family-group taxa. The intuitively reconstructed phylogeny of 75 extant genera failed to recover those taxa or their relationships. Parsimony analyses of various datasets all yielded strong polyphyly and chaotic relationships. In conclusion, the forewing of cockroaches is not a universally competent character system. The underlying causes are the complicated nature of veins and the limitations of current analytical techniques. The uncertainty in forewing-based taxonomy and systematics has been underestimated in the literature. Forewing-based fossil taxa warrant re-evaluation; some of them are herein deemed nomina dubia in their current state.

The postcranial skeleton of Teyujagua paradoxa (Reptilia: Archosauromorpha) from the early Triassic of South America.

Teyujagua paradoxa is a remarkable early archosauromorph from the Lower Triassic Sanga do Cabral Formation, Brazil. The species was originally described from an almost complete skull and a few associated cervical vertebrae, and no further postcranial elements were known at that time. Additional fieldwork in the Sanga do Cabral Formation, however, was successful in recovering a fairly complete postcranial skeleton attributable to the holotype. Here, we describe this new postcranial material, which is composed of cervical, dorsal, sacral and caudal vertebrae, limbs, pectoral and pelvic girdles, ribs, and gastralia. The description of its postcranial skeleton makes T. paradoxa one of the best-known early-diverging archosauromorphs. The cladistic analysis performed after the scoring of postcranial data recovered T. paradoxa in the same position initially described, close to the node that defines the Archosauriformes. Teyujagua paradoxa shares morphological features with representatives of early-diverging archosauromorphs and archosauriforms, with certain traits demonstrating a mosaic of plesiomorphic and apomorphic character states. We also performed partitioned morphospace and disparity analysis to elucidate the morphological disparity and evolutionary patterns among archosauromorphs. Teyujagua paradoxa occupies a notable position, suggesting an intermediate morphology between early archosauromorphs and proterosuchids. Disparity estimates highlighted Pseudosuchia and Avemetatarsalia as having the highest median disparity, reflecting their diverse cranial and postcranial morphologies, respectively. These findings offer valuable insights into archosauromorph macroevolution and adaptation.

Quantifying the Uncertainty of Imputed Demographic Disparity Estimates: The Dual-Bootstrap

Quantifying the Uncertainty of Imputed Demographic Disparity Estimates: The Dual-Bootstrap

An Interplay between Health Equity and Disparity in Cognitive Decline

AbstractBackgroundCognitive decline may be an early indicator of Alzheimer diseases and related dementia (ADRD). To identify the intervention targets that ameliorate disparities in ADRD, it is critical to employ disparity measures that account for health equity. We aimed to investigate disparity estimates in cognitive decline built on three health equity measures.MethodThe study sample (n = 451) was derived from the bi‐ethnic community‐dwelling San Antonio Longitudinal Study of Aging (SALSA). The latent variable mixture model classified cognitive function trajectory strata based on four repeated measures of the Mini Mental Status Examination (MMSE) over 8.5±1.3 years of follow‐up. Three risk adjustments were employed to optimize equity in social determinants of health (SDoH: income, education, social interactions) between Hispanics and non‐Hispanics. 1) Average constrained regression adjustment minimized the impact of SDoH for Hispanics. 2) Covariance constrained adjustment derived estimates with a near independence between the residuals with ethnicity. 3) Net compensation penalized regression imposed a penalty for excess lower MMSE among Hispanics. Inverse propensity score weighting (IPSW) estimates of disparity further adjusted away confounding due to covariates (demographics, disease burdens) between ethnic groups.ResultTable 1 contrasted baseline characteristics of the study sample between Hispanics and non‐Hispanics. Two MMSE trajectory strata were identified: cognitive decline stratum (38%) with mean MMSE decreased from 23.9 to 21.2; and stable cognitive function stratum (62%) with mean MMSE around 28 throughout. Average constrained and covariance constrained risk adjustments attained optimal SDoH equity across three fairness metrics: predictive ratios, fair covariance, and net compensation (Table 2). Ethnic disparity in cognitive decline (Table 3) was strong conditioned on SDoH equity alone (log(aOR) = 4.27, p<0.001). Balancing covariates alone via IPSW without equity in SDoH also fell short for parity (aOR = 1.16, p<0.001). Parity in cognitive decline could be achievable under both SDoH equity and balanced covariates between ethnic groups (aOR = 0.85, p = 0.21).ConclusionOur study suggested that integrating equity metrices in ADRD disparity research could guide the paths to improve equity and reduce disparity. The field would be better served by identifying the causal drivers of disparities through the advancement in equity metrices and mediation analyses that delineate intricate interactions between SDoH and ADRD risk factors.

Cross-sectional population-based estimates of a rural-urban disparity in prevalence of long COVID among Michigan adults with polymerase chain reaction-confirmed COVID-19, 2020-2022.

To (1) assess whether residential rurality/urbanicity was associated with the prevalence of 30- or 90-day long COVID, and (2) evaluate whether differences in long COVID risk factors might explain this potential disparity. We used data from the Michigan COVID-19 Recovery Surveillance Study, a population-based probability sample of adults with COVID-19 (n = 4,937). We measured residential rurality/urbanicity using dichotomized Rural-Urban Commuting Area codes (metropolitan, nonmetropolitan). We considered outcomes of 30-day long COVID (illness duration ≥30 days) and 90-day long COVID (illness duration ≥90 days). Using Poisson regression, we estimated unadjusted prevalence ratios (PRs) to compare 30- and 90-day long COVID between metropolitan and nonmetropolitan respondents. Then, we adjusted our model to account for differences between groups in long COVID risk factors (age, sex, acute COVID-19 severity, vaccination status, race and ethnicity, socioeconomic status, health care access, SARS-CoV-2 variant, and pre-existing conditions). We estimated associations for the full study period (Jan 1, 2020-May 31, 2022), the pre-vaccine era (before April 5, 2021), and the vaccine era (after April 5, 2021). Compared to metropolitan adults, the prevalence of 30-day long COVID was 15% higher (PR = 1.15 [95% CI: 1.03, 1.29]), and the prevalence of 90-day long COVID was 27% higher (PR = 1.27 [95% CI: 1.09, 1.49]) among nonmetropolitan adults. Adjusting for long COVID risk factors did not reduce disparity estimates in the pre-vaccine era but halved estimates in the vaccine era. Our findings provide evidence of a rural-urban disparity in long COVID and suggest that the factors contributing to this disparity changed over time as the sociopolitical context of the pandemic evolved and COVID-19 vaccines were introduced.

Mitigating invalid and mischievous survey responses: A registered report examining risk disparities between heterosexual and lesbian, gay, bisexual, or questioning youth.

In a Registered Report, the authors propose a new survey-bias-mitigation method-incorporating inverse probability weighting via boosted regression-to better understand lesbian, gay, bisexual, or questioning (LGBQ)-heterosexual youth risk disparities. This method is tested using the 2019 US Centers for Disease Control and Prevention-collected Youth Risk Behavior Survey (YRBS) national data, which contain 12,847 observations (ages 12-18 [M = 16, SD = 1.25]; 49.1% male [8.7% LGBQ] and 50.9% female [22.4% LGBQ]; nationally representative regarding race and ethnicity). Looking across 44 outcomes, the authors found that the YRBS contains responses that are potentially biased against LGBQ youth in systematic ways, inflating perceived risk for this group in some outcomes. This potential bias is more pronounced among reported males than among reported females, and it is more pronounced for low-incidence outcomes. For example, the steroid-use disparity estimate among reported males reduced by 67%, while the reduction in bullying victimization was small and not statistically significant. The authors discuss robustness of results, the new method, and data policy implications.

Partially Recursively Induced Structured Moderation (PRISM) for modeling racial differences in endometrial cancer survival.

Health disparities are driven by a complex interplay of determinants operating across multiple levels of influence. However, while recognized conceptually, much disparities research fails to capture this inherent complexity in study focus and/or design; little of such work accounts for the interplay across the multiple levels of influence from structural (contextual) to biological or clinical. We developed a novel modeling framework that addresses these challenges and provides new insights. We used data from the Florida Cancer Data System on endometrial cancer patients and geocoded-derived social determinants of health to demonstrate the applicability of a new modeling paradigm we term PRISM regression. PRISM is a new highly interpretable tree-based modeling framework that allows for automatic discovery of potentially non-linear hierarchical interactions between health determinants at multiple levels and differences in survival outcomes between groups of interest, including through a new specific area-level disparity estimate (SPADE) incorporating these multilevel influences. PRISM demonstrates that hierarchical influences on racial disparity in endometrial cancer survival appear to be statistically relevant and that these better predict survival differences than only using individual level determinants. The interpretability of the models allows more careful inspection of the nature of these hierarchical effects on disparity. Additionally, SPADE estimates show distinct geographical patterns across census tracts in Florida. PRISM can provide a powerful new modeling framework with which to better understand racial disparities in cancer survival.

A novel factor graph-based optimization technique for stereo correspondence estimation

Dense disparities among multiple views are essential for estimating the 3D architecture of a scene based on the geometrical relationship between the scene and the views or cameras. Scenes with larger extents of homogeneous textures, differing scene illumination among the multiple views and with occluding objects affect the accuracy of the estimated disparities. Markov random fields based methods for disparity estimation address these limitations using spatial dependencies among the observations and among the disparity estimates. These methods, however, are limited by spatially fixed and smaller neighborhood systems or cliques. Recent learning-based methods generate rich set of stereo features for generating cost volume and estimating disparity. In this work, we present a new factor graph-based probabilistic graphical model for disparity estimation that allows a larger and a spatially variable neighborhood structure determined based on the local scene characteristics. Our algorithm improves the accuracy of disparity estimates in stereo image pairs with varying texture and illumination characteristics by enforcing spatial dependencies among scene characteristics as well as among disparity estimates. We evaluated our method using the Middlebury benchmark stereo datasets and the Middlebury evaluation dataset version 3.0 and compared its performance with recent state-of-the-art disparity estimation algorithms. Our factor graph-based algorithm provided disparity estimates with higher accuracy when compared to the recent non-learning- and learning-based disparity estimation algorithms. The factor graph formulation can be used for obtaining maximum a posteriori estimates from models or optimization problems with complex dependency structure among hidden variables. The strategies of using a priori distributions with shorter support and spatial dependencies were useful for reducing the computational cost and improving message convergence in the model. The factor-graph algorithm is also useful for other dense estimation problems such as optical flow estimation.

National and Intraurban Air Pollution Exposure Disparity Estimates in the United States: Impact of Data-Aggregation Spatial Scale.

Air pollution exposure disparities by race/ethnicity and socioeconomic status have been analyzed using data aggregated at various spatial scales. Our research question is this: To what extent does the spatial scale of data aggregation impact the estimated exposure disparities? We compared disparities calculated using data spatially aggregated at five administrative scales (state, county, census tract, census block group, census block) in the contiguous United States in 2010. Specifically, for each of the five spatial scales, we calculated national and intraurban disparities in exposure to fine particles (PM2.5) and nitrogen dioxide (NO2) by race/ethnicity and socioeconomic characteristics using census demographic data and an empirical statistical air pollution model aggregated at that scale. We found, for both pollutants, that national disparity estimates based on state and county scale data often substantially underestimated those estimated using tract and finer scales; in contrast, national disparity estimates were generally consistent using tract, block group, and block scale data. Similarly, intraurban disparity estimates based on tract and finer scale data were generally well correlated for both pollutants across urban areas, although in some cases intraurban disparity estimates were substantially different, with tract scale data more frequently leading to underestimates of disparities compared to finer scale analyses.

Abstract 3672: Area-level drivers of the breast cancer mortality race disparity in Georgia

Abstract In the US, breast cancer mortality is 40% greater among Black than White women. The mortality disparity varies geographically and has persisted despite improvements in survival. Georgia is one of six states where breast cancer recently surpassed lung cancer as the leading cause of cancer death among Black but not White women, is a large and diverse state, and thus, an ideal setting to study race disparities. Our work in the Atlanta Metro area has shown that the mortality disparity is pronounced among patients with clinically favorable subtypes, low Oncotype DX® recurrence scores, guideline-concordant care, and similar access to surgical facilities. The disparity also persists among patients with private insurance and those living in areas of low poverty, suggesting clinical and sociodemographic differences do not fully account for the observed disparity. Recent evidence suggests area-level factors may contribute to the disparity. Our group reported that neighborhood-level redlining (i.e., mortgage denial based on place) is associated with an increase in breast cancer mortality while neighborhood-level lending bias (i.e., mortgage denial based on race) is associated with a decrease in breast cancer mortality. In Atlanta, Black patients are more likely to live in redlined areas and less likely to live in areas with lending bias, than White patients, further suggesting place is an important driver of disparities. We investigated the role of other area-level factors on race-specific mortality in Georgia. Race-specific standardized mortality ratios were computed for each county in Georgia. Observed breast cancer deaths among non-Hispanic White (NHW) and non-Hispanic Black (NHB) women within 5 years of a stage I-III breast cancer diagnosis (2005-2013) were obtained from the Georgia Cancer Registry. County-level characteristics were derived from the American Community Survey and Georgia Department of Public Health. A Bayesian model-based approach was used to stabilize local estimates and estimate associations with area-level factors. Smoothed estimates of relative disparity ranged from 1.2 to 1.7, with pronounced disparity for counties surrounding the Atlanta and Savannah metropolitan areas and least pronounced disparity in Metro Albany counties. Among NHB women, living in a rural (versus non-rural) county was associated with a nearly 20% increase in breast cancer mortality (RR=1.19, 95% credible interval (CI): 0.96, 1.48); an association not observed in NHW patients (RR=0.93, 95% CI: 0.79, 1.07). In contrast, a 10-point increase in the percent of adults with ≤high school education was associated with a 10% increase (95% CI: 1.05, 1.17) in mortality for NHW patients only. The race disparity in Georgia is not evenly distributed across counties. We identified area-level characteristics associated with race-specific mortality. Identifying other area-level drivers of breast cancer mortality is an important area of inquiry. Citation Format: Rebecca Nash, Lindsay J. Collin, Jeffrey Switchenko, Lauren E. McCullough. Area-level drivers of the breast cancer mortality race disparity in Georgia [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 3672.

Sex disparities in enrollment and reporting of outcomes by sex in contemporary clinical trials of atrial fibrillation.

BackgroundUnderrepresentation of females in randomized controlled trials (RCTs) limits generalizability and quality of the evidence guiding treatment of females. This study aimed to measure the sex disparities in participants' recruitment in RCTs of atrial fibrillation (AF) and determine associated factors, and to describe the frequency of outcomes reported by sex.MethodsMEDLINE was searched to identify RCTs of AF published between January 1, 2011, and November 20, 2021, in 12 top‐tier journals. We measured the enrollment of females using the enrollment disparity difference (EDD) which is the difference between the proportion of females in the trial and the proportion of females with AF in the underlying general population (obtained from the Global Burden of Disease). Random‐effects meta‐analyses of the EDD were performed, and multivariable meta‐regression was used to explore factors associated with disparity estimates. We also determined the proportion of trials that included sex‐stratified results.ResultsOut of 1133 records screened, 142 trials were included, reporting on a total of 133 532 participants. The random‐effects summary EDD was −0.125 (95% confidence interval [CI] = −0.143 to −0.108), indicating that females were under‐enrolled by 12.5 percentage points. Female enrollment was higher in trials with higher sample size (<250 vs. >750, adjusted odds ratio [aOR] 1.065, 95% CI: 1.008–1.125), higher mean participants' age (aOR: 1.006, 95% CI: 1.002–1.009), and lower in trials conducted in North America compared to Europe (aOR: 0.945, 95% CI: 0.898–0.995). Only 36 trials (25.4%) reported outcomes by sex, and of these 29 (80.6%) performed statistical testing of the sex‐by‐treatment interaction.ConclusionFemales remain substantially less represented in RCTs of AF, and sex‐stratified reporting of primary outcomes is infrequent. These findings call for urgent action to improve sex equity in enrollment and sex‐stratified outcomes' reporting in RCTs of AF.

Edge supervision and multi-scale cost volume for stereo matching

Recently, methods based on Convolutional Neural Network have achieved huge progress in stereo matching. However, it is still difficult to find accurate matching points in inherently ill-posed regions (e.g., weak texture areas and around object edges), in which the accuracy of disparity estimate can be improved by the corresponding geometric constraints. To tackle this problem, we innovatively generate the depth ground-truth boundary dataset by mining the instance segmentation and semantic segmentation datasets and propose RDNet, which incorporates edge cues into stereo matching. The network learns geometric information through a separate processing branch edge stream, which can process feature information in parallel with the stereo stream. The edge stream removes noise and only focuses on processing the relevant boundary information. Besides, we introduce a multi-scale cost volume in hierarchical cost aggregation to enlarge the receptive fields and capture structural and global representations that can significantly improve the ability of scene understanding and disparity estimation accuracy. Moreover, a disparity refinement network with several dilated convolutions is applied to further improve the accuracy of the final disparity estimation. The proposed method is evaluated on Sceneflow, KITTI 2015 and KITTI 2012 benchmark datasets, and the qualitative and quantitative results demonstrate that the proposed RDNet significantly achieves the state-of-the-art stereo matching performance.

Black Patients With Cirrhosis Have Higher Mortality and Lower Transplant Rates: Results From a Metropolitan Cohort Study.

Estimates of racial disparity in cirrhosis have been limited by lack of large-scale, longitudinal data, which track patients from diagnosis to death and/or transplant. We analyzed a large, metropolitan, population-based electronic health record data set from seven large health systems linked to the state death registry and the national transplant database. Multivariate competing risk analyses, adjusted for sex, age, insurance status, Elixhauser score, etiology of cirrhosis, HCC, portal hypertensive complication, and Model for End-Stage Liver Disease-Sodium (MELD-Na), examined the relationship between race, transplant, and cause of death as defined by blinded death certificate review. During the study period, 11,277 patients met inclusion criteria, of whom 2,498 (22.2%) identified as Black. Compared to White patients, Black patients had similar age, sex, MELD-Na, and proportion of alcohol-associated liver disease, but higher comorbidity burden, lower rates of private insurance, and lower rates of portal hypertensive complications. Compared to White patients, Black patients had the highest rate all-cause mortality and non-liver-related death and were less likely to be listed or transplanted (P<0.001 for all). In multivariate competing risk analysis, Black patients had a 26% increased hazard of liver-related death (subdistribution HR, 1.26; 95% CI, [1.15-1.38]; P<0.001). Black patients with cirrhosis have discordant outcomes. Further research is needed to determine how to address these real disparities in the field of hepatology.

Equal Protection by Race with Stop and Frisk: a Risk-Adjusted Disparity (RAD) Index for Balanced Policing

Research QuestionCan racial equity in crime and policing be measured with the use of a Risk-Adjusted Disparity (RAD) Index of the degree to which policing across racial categories is “balanced” in its ratios of preventive police actions per 100 serious crimes committed against members of each racial category?DataOffice of National Statistics (ONS) reports on crime and policing in England and Wales, and Dorset Police data on violent crime victimization and stop-search by race of suspect across the 452 Lower-Layer Super-Output Levels in Dorset.MethodsWe conceptualize the problem of equal protection under law as fundamentally protecting the lives and liberties of each citizen from criminal harms, as well as from disproportionately intrusive policing. We combine these dimensions into a single metric that defines proportionality of policing in relation to risk of violent crime victimization, such that whatever intrusion on liberty is applied for the aim of protection can be equalized across racial groups.FindingsThe use of a Risk-Adjusted Disparity (RAD) Index to measure reliably the equality of police intrusions across racial groups based on victimization rates can be illustrated by adjusting for homicide. In the past decade, the population-based disparity rate shows that Blacks are stopped by police nine times more often than whites. When that rate is adjusted for the differential risk of homicide in the two groups, the disparity estimate drops from 800% to 58%. Other changes of major magnitude result from using the RAD Index.ConclusionsWe conclude that an index of proactive policing using victimizations by race is more likely to lead to equal protection of law than a residential population-based metric of proactive police actions, as is commonly used in official reporting. A victim-based, Risk-Adjusted (RAD) Index for measuring racial disparity might focus police efforts on the 5% of local areas where serious violence is concentrated, and deflect stops away from the vast majority of areas that have little serious crime.

COVID-19-related state-wise racial and ethnic disparities across the USA: an observational study based on publicly available data from The COVID Tracking Project

ObjectiveTo evaluate COVID-19 infection and mortality disparities in ethnic and racial subgroups in a state-wise manner across the USA.MethodsPublicly available data from The COVID Tracking Project at The Atlantic were...

Stereopagnosia: Fooling Stereo Networks with Adversarial Perturbations

We study the effect of adversarial perturbations of images on the estimates of disparity by deep learning models trained for stereo. We show that imperceptible additive perturbations can significantly alter the disparity map, and correspondingly the perceived geometry of the scene. These perturbations not only affect the specific model they are crafted for, but transfer to models with different architecture, trained with different loss functions. We show that, when used for adversarial data augmentation, our perturbations result in trained models that are more robust, without sacrificing overall accuracy of the model. This is unlike what has been observed in image classification, where adding the perturbed images to the training set makes the model less vulnerable to adversarial perturbations, but to the detriment of overall accuracy. We test our method using the most recent stereo networks and evaluate their performance on public benchmark datasets.

Evolution Towards Fossoriality and Morphological Convergence in the Skull of Spalacidae and Bathyergidae (Rodentia)

Rodents show a wide range of anatomical, physiological, and behavioral adaptations to life underground. Cranial and postcranial bone morphologies are deeply impacted by the modes of digging, which can involve either incisors or claws. However, the morphological variation of these elements still needs to be accurately quantified to assess the degree of specializations of the fossorial rodent families in regards to their respective evolution. Here, we focus on the morpho-functional characteristics of the masticatory apparatus in two families of subterranean rodents, the Spalacidae and Bathyergidae. We quantify skull shape in five spalacid genera using geometric morphometric methods, as well as biomechanical estimates for adductor muscles, which are compared with data previously published on bathyergids. We show that skull shape of spalacids has a greater disparity and lower biomechanical estimates than bathyergids, in which the fossorial activity, notably chisel-tooth digging, more significantly impacted the evolution of the skull. Among spalacids, Spalax shows the most extreme specializations to life underground and displays the highest number of morphological convergences with chisel-tooth digging bathyergids, especially regarding its cranial shape and high biomechanical estimate for the temporalis muscle. Fewer morphological convergences were observed between other spalacids and bathyergids. Different evolutionary histories can potentially explain discrepancies observed between the two families, the first bathyergid morphological adaptations to fossorial life being much older than those of spalacids.

Unsupervised Occlusion-Aware Stereo Matching With Directed Disparity Smoothing

When handling occlusion in unsupervised stereo matching, existing methods tend to neglect the supportive role of occlusion and to perform inappropriate disparity smoothing around the occlusion. To address these problems, we propose an occlusion-aware stereo network that contains a specific module to first estimate occlusion as an additional depth cue. In the occlusion inference module, a pixel is classified with a three-category label based on whether an area is occluded by an object on the left, occluded by an object on the right, or unoccluded. After the occluders are detected, we introduce a directed disparity smoothing loss that allows valid disparity estimates to be propagated to fill the occluded region, while ambiguous matches in the occluded region do not affect other regions. Disparity and occlusion are trained alternately in an unsupervised manner with detached backpropagation to enable the directed smoothness. Experiments show that our method achieves 3-pixel threshold error rates of 6.51% and 5.69% on the KITTI 2015 and KITTI 2012 validation sets, state-of-the-art results among unsupervised learning networks at the time of submission.