Recently, significant progress has been made applying machine learning to the problem of table structure inference and extraction from unstructured documents. However, one of the greatest challenges remains the creation of datasets with complete, unambiguous ground truth at scale. To address this, we develop a new, more comprehensive dataset for table extraction, called PubTables-1M. PubTables-1M contains nearly one million tables from scientific articles, supports multiple input modalities, and contains detailed header and location information for table structures, making it useful for a wide variety of modeling approaches. It also addresses a significant source of ground truth inconsistency observed in prior datasets called oversegmentation, using a novel canonicalization procedure. We demonstrate that these improvements lead to a significant increase in training performance and a more reliable estimate of model performance at evaluation for table structure recognition. Further, we show that transformer-based object detection models trained on PubTables-1M produce excellent results for all three tasks of detection, structure recognition, and functional analysis without the need for any special customization for these tasks. Data and code will be released at https://github.com/microsoft/table-transformer.

While the Choquet integral (Chi) is a powerful parametric nonlinear aggregation function, it has limited scope and is not a universal function generator. Herein, we focus on a class of problems that are outside the scope of a single Chi. Namely, we are interested in tasks where different subsets of inputs require different Chls. Herein, a genetic program (GF) is used to extend the Chi, referred to as GpChI hereafter, specifically in terms of compositions of Chls and/or arithmetic combinations of Chls. An algorithm is put forth to leam the different GP Chls via genetic algorithm (GA) optimization. Synthetic experiments demonstrate GpChI in a controlled fashion, i.e., we know the answer and can compare what is learned to the truth. Real-world experiments are also provided for the mult-sensor fusion of electromagnetic induction (EMI) and ground penetrating radar (GPR) for explosive hazard detection. Our mutli-sensor fusion experiments show that there is utility in changing aggregation strategy per different subsets of inputs (sensors or algorithms) and fusing those results.

In a standard target detection approach, data is collected, points of interest called alarms are identified, and detection algorithms determine the confidence that a target is present at each point. Receiver operating characteristic (ROC) curves can be used to evaluate the performance of each detector and choose operating thresholds. The use of multiple sensors can improve the probability of detection of a diverse set of targets. It is difficult to properly assess the performance of a system of detectors and choose the best joint set of operating thresholds if confidence values from different detectors do not compare meaningfully. Fusion methods can be used to improve the joint performance of a set of detectors. However, in the case where different detectors do not operate on the same points of interest, typical fusion methods cannot be used to improve the binary decisions on individual alarms. In this paper, we propose a new fusion method that maps the confidence outputs from different detectors to a shared range where they compare meaningfully, and optimizes the joint performance of multiple detectors even when their alarm sets are disjoint. Our method uses dynamic programming to monotonically map the confidence output from each detector onto a shared range in such a way that we maximize the area-under-the-curve (AUC) of the ROC curve corresponding to the joint set of alarms. This joint ROC curve can be used to determine the operational thresholds for each individual detector to maximize their joint performance.

Identifying the ground surface in ground-penetrating radar (GPR) data is useful and can be done efficiently and accurately using the Viterbi algorithm. This involves representing the radar image as a trellis graph and solving for the optimal path. To identify multiple layer boundaries in a radar image in this manner, it is necessary to find multiple disjoint paths through the trellis. Two main types of algorithms currently exist that find the k best disjoint paths whose aggregate sum is minimized. However, this criterion has drawbacks. Instead, we propose a novel criterion for choosing multiple disjoint paths in a trellis that we call the reciprocal pointer chain. This criterion has both a nice intuitive and theoretical justification, and leads to an algorithm with better qualitative results and significantly lower computational complexity than any of the methods previously proposed.