Thomas Manzini

Thomas Manzini, Priyankari Perali, Raisa Karnik et al. · microsoft-research

This document presents the Center for Robot Assisted Search And Rescue - Uncrewed Aerial Systems - Disaster Response Overhead Inspection Dataset (CRASAR-U-DROIDs) for building damage assessment and spatial alignment collected from small uncrewed aerial systems (sUAS) geospatial imagery. This dataset is motivated by the increasing use of sUAS in disaster response and the lack of previous work in utilizing high-resolution geospatial sUAS imagery for machine learning and computer vision models, the lack of alignment with operational use cases, and with hopes of enabling further investigations between sUAS and satellite imagery. The CRASAR-U-DRIODs dataset consists of fifty-two (52) orthomosaics from ten (10) federally declared disasters (Hurricane Ian, Hurricane Ida, Hurricane Harvey, Hurricane Idalia, Hurricane Laura, Hurricane Michael, Musset Bayou Fire, Mayfield Tornado, Kilauea Eruption, and Champlain Towers Collapse) spanning 67.98 square kilometers (26.245 square miles), containing 21,716 building polygons and damage labels, and 7,880 adjustment annotations. The imagery was tiled and presented in conjunction with overlaid building polygons to a pool of 130 annotators who provided human judgments of damage according to the Joint Damage Scale. These annotations were then reviewed via a two-stage review process in which building polygon damage labels were first reviewed individually and then again by committee. Additionally, the building polygons have been aligned spatially to precisely overlap with the imagery to enable more performant machine learning models to be trained. It appears that CRASAR-U-DRIODs is the largest labeled dataset of sUAS orthomosaic imagery.

Thomas Manzini, Robin Murphy · microsoft-research

This paper details the challenges in applying two computer vision systems, an EfficientDET supervised learning model and the unsupervised RX spectral classifier, to 98.9 GB of drone imagery from the Wu-Murad wilderness search and rescue (WSAR) effort in Japan and identifies 3 directions for future research. There have been at least 19 proposed approaches and 3 datasets aimed at locating missing persons in drone imagery, but only 3 approaches (2 unsupervised and 1 of an unknown structure) are referenced in the literature as having been used in an actual WSAR operation. Of these proposed approaches, the EfficientDET architecture and the unsupervised spectral RX classifier were selected as the most appropriate for this setting. The EfficientDET model was applied to the HERIDAL dataset and despite achieving performance that is statistically equivalent to the state-of-the-art, the model fails to translate to the real world in terms of false positives (e.g., identifying tree limbs and rocks as people), and false negatives (e.g., failing to identify members of the search team). The poor results in practice for algorithms that showed good results on datasets suggest 3 areas of future research: more realistic datasets for wilderness SAR, computer vision models that are capable of seamlessly handling the variety of imagery that can be collected during actual WSAR operations, and better alignment on performance measures.

Robin Murphy, Thomas Manzini · microsoft-research

This paper describes gaps in acquisition of drone imagery that impair the use with computer vision/machine learning (CV/ML) models and makes five recommendations to maximize image suitability for CV/ML post-processing. It describes a notional work process for the use of drones in wilderness search and rescue incidents. The large volume of data from the wide area search phase offers the greatest opportunity for CV/ML techniques because of the large number of images that would otherwise have to be manually inspected. The 2023 Wu-Murad search in Japan, one of the largest missing person searches conducted in that area, serves as a case study. Although drone teams conducting wide area searches may not know in advance if the data they collect is going to be used for CV/ML post-processing, there are data collection procedures that can improve the search in general with automated collection software. If the drone teams do expect to use CV/ML, then they can exploit knowledge about the model to further optimize flights.

Thomas Manzini, Priyankari Perali, Robin R. Murphy

This paper presents the first AI/ML system for automating building damage assessment in uncrewed aerial systems (sUAS) imagery to be deployed operationally during federally declared disasters (Hurricanes Debby and Helene). In response to major disasters, sUAS teams are dispatched to collect imagery of the affected areas to assess damage; however, at recent disasters, teams collectively delivered between 47GB and 369GB of imagery per day, representing more imagery than can reasonably be transmitted or interpreted by subject matter experts in the disaster scene, thus delaying response efforts. To alleviate this data avalanche encountered in practice, computer vision and machine learning techniques are necessary. While prior work has been deployed to automatically assess damage in satellite imagery, there is no current state of practice for sUAS-based damage assessment systems, as all known work has been confined to academic settings. This work establishes the state of practice via the development and deployment of models for building damage assessment with sUAS imagery. The model development involved training on the largest known dataset of post-disaster sUAS aerial imagery, containing 21,716 building damage labels, and the operational training of 91 disaster practitioners. The best performing model was deployed during the responses to Hurricanes Debby and Helene, where it assessed a combined 415 buildings in approximately 18 minutes. This work contributes documentation of the actual use of AI/ML for damage assessment during a disaster and lessons learned to the benefit of the AI/ML research and user communities.

Thomas Manzini, Priyankari Perali, Jayesh Tripathi et al. · microsoft-research

This paper audits damage labels derived from coincident satellite and drone aerial imagery for 15,814 buildings across Hurricanes Ian, Michael, and Harvey, finding 29.02% label disagreement and significantly different distributions between the two sources, which presents risks and potential harms during the deployment of machine learning damage assessment systems. Currently, there is no known study of label agreement between drone and satellite imagery for building damage assessment. The only prior work that could be used to infer if such imagery-derived labels agree is limited by differing damage label schemas, misaligned building locations, and low data quantities. This work overcomes these limitations by comparing damage labels using the same damage label schemas and building locations from three hurricanes, with the 15,814 buildings representing 19.05 times more buildings considered than the most relevant prior work. The analysis finds satellite-derived labels significantly under-report damage by at least 20.43% compared to drone-derived labels (p<1.2x10^-117), and satellite- and drone-derived labels represent significantly different distributions (p<5.1x10^-175). This indicates that computer vision and machine learning (CV/ML) models trained on at least one of these distributions will misrepresent actual conditions, as the differing satellite and drone-derived distributions cannot simultaneously represent the distribution of actual conditions in a scene. This potential misrepresentation poses ethical risks and potential societal harm if not managed. To reduce the risk of future societal harms, this paper offers four recommendations to improve reliability and transparency to decisio-makers when deploying CV/ML damage assessment systems in practice

Thomas Manzini, Priyankari Perali, Raisa Karnik et al. · microsoft-research

This work presents the first quantitative study of alignment errors between small uncrewed aerial systems (sUAS) georectified imagery and a priori building polygons and finds that alignment errors are non-uniform and irregular, which negatively impacts field robotics systems and human-robot interfaces that rely on geospatial information. There are no efforts that have considered the alignment of a priori spatial data with georectified sUAS imagery, possibly because straight-forward linear transformations often remedy any misalignment in satellite imagery. However, an attempt to develop machine learning models for an sUAS field robotics system for disaster response from nine wide-area disasters using the CRASAR-U-DROIDs dataset uncovered serious translational alignment errors. The analysis considered 21,608 building polygons in 51 orthomosaic images, covering 16787.2 Acres (26.23 square miles), and 7,880 adjustment annotations, averaging 75.36 pixels and an average intersection over union of 0.65. Further analysis found no uniformity among the angle and distance metrics of the building polygon alignments, presenting an average circular variance of 0.28 and an average distance variance of 0.45 pixels2, making it impossible to use the linear transform used to align satellite imagery. The study's primary contribution is alerting field robotics and human-robot interaction (HRI) communities to the problem of spatial alignment and that a new method will be needed to automate and communicate the alignment of spatial data in sUAS georectified imagery. This paper also contributes a description of the updated CRASAR-U-DROIDs dataset of sUAS imagery, which contains building polygons and human-curated corrections to spatial misalignment for further research in field robotics and HRI.

Thomas Manzini, Priyankari Perali, Raisa Karnik et al.

This paper presents the largest known benchmark dataset for road damage assessment and road alignment, and provides 18 baseline models trained on the CRASAR-U-DRIODs dataset's post-disaster small uncrewed aerial systems (sUAS) imagery from 10 federally declared disasters, addressing three challenges within prior post-disaster road damage assessment datasets. While prior disaster road damage assessment datasets exist, there is no current state of practice, as prior public datasets have either been small-scale or reliant on low-resolution imagery insufficient for detecting phenomena of interest to emergency managers. Further, while machine learning (ML) systems have been developed for this task previously, none are known to have been operationally validated. These limitations are overcome in this work through the labeling of 657.25km of roads according to a 10-class labeling schema, followed by training and deploying ML models during the operational response to Hurricanes Debby and Helene in 2024. Motivated by observed road line misalignment in practice, 9,184 road line adjustments were provided for spatial alignment of a priori road lines, as it was found that when the 18 baseline models are deployed against real-world misaligned road lines, model performance degraded on average by 5.596\% Macro IoU. If spatial alignment is not considered, approximately 8\% (11km) of adverse conditions on road lines will be labeled incorrectly, with approximately 9\% (59km) of road lines misaligned off the actual road. These dynamics are gaps that should be addressed by the ML, CV, and robotics communities to enable more effective and informed decision-making during disasters.

Thomas Manzini, Yao Chong Lim, Yulia Tsvetkov et al.

Online texts -- across genres, registers, domains, and styles -- are riddled with human stereotypes, expressed in overt or subtle ways. Word embeddings, trained on these texts, perpetuate and amplify these stereotypes, and propagate biases to machine learning models that use word embeddings as features. In this work, we propose a method to debias word embeddings in multiclass settings such as race and religion, extending the work of (Bolukbasi et al., 2016) from the binary setting, such as binary gender. Next, we propose a novel methodology for the evaluation of multiclass debiasing. We demonstrate that our multiclass debiasing is robust and maintains the efficacy in standard NLP tasks.

Hai Pham, Paul Pu Liang, Thomas Manzini et al.

Multimodal sentiment analysis is a core research area that studies speaker sentiment expressed from the language, visual, and acoustic modalities. The central challenge in multimodal learning involves inferring joint representations that can process and relate information from these modalities. However, existing work learns joint representations by requiring all modalities as input and as a result, the learned representations may be sensitive to noisy or missing modalities at test time. With the recent success of sequence to sequence (Seq2Seq) models in machine translation, there is an opportunity to explore new ways of learning joint representations that may not require all input modalities at test time. In this paper, we propose a method to learn robust joint representations by translating between modalities. Our method is based on the key insight that translation from a source to a target modality provides a method of learning joint representations using only the source modality as input. We augment modality translations with a cycle consistency loss to ensure that our joint representations retain maximal information from all modalities. Once our translation model is trained with paired multimodal data, we only need data from the source modality at test time for final sentiment prediction. This ensures that our model remains robust from perturbations or missing information in the other modalities. We train our model with a coupled translation-prediction objective and it achieves new state-of-the-art results on multimodal sentiment analysis datasets: CMU-MOSI, ICT-MMMO, and YouTube. Additional experiments show that our model learns increasingly discriminative joint representations with more input modalities while maintaining robustness to missing or perturbed modalities.

Hai Pham, Thomas Manzini, Paul Pu Liang et al.

Multimodal machine learning is a core research area spanning the language, visual and acoustic modalities. The central challenge in multimodal learning involves learning representations that can process and relate information from multiple modalities. In this paper, we propose two methods for unsupervised learning of joint multimodal representations using sequence to sequence (Seq2Seq) methods: a \textit{Seq2Seq Modality Translation Model} and a \textit{Hierarchical Seq2Seq Modality Translation Model}. We also explore multiple different variations on the multimodal inputs and outputs of these seq2seq models. Our experiments on multimodal sentiment analysis using the CMU-MOSI dataset indicate that our methods learn informative multimodal representations that outperform the baselines and achieve improved performance on multimodal sentiment analysis, specifically in the Bimodal case where our model is able to improve F1 Score by twelve points. We also discuss future directions for multimodal Seq2Seq methods.