On-the-Fly Performance Evaluation of Large-Scale Fiber Track- ing

Large-scale fiber tracking in the images serial-sectioned from material samples is a critical step to analyze physical properties of continuous fiber reinforced composite materials. In serial-section imaging, increasing the sampling sparsity, i.e., the inter-slice intervals, can lead to significant speedups in data collection. However, increasing the sampling sparsity leads to difficulties in tracking large-scale crowded and similar-appearance fibers through the serial-section slices. One way to address this issue is to dynamically adjust the sampling rate by balancing the tracking accuracy with the data collection time. For this purpose, it is necessary to develop methods for estimating the tracking accuracy on the fly, i.e., immediately after tracking is updated on a new serial-section slice. Typical tracking accuracy metrics require ground truths, which are usually constructed by human annotations and unavailable on the fly. In this paper, we present a new approach to evaluate the performance of online largescale fiber tracking without involving the ground truth. Specifically, we explore the local spatial consistency of the fibers between adjacent slices and define a new performance-evaluation metric based on this spatial consistency. A set of experiments on real composite-material images are conducted to illustrate the effectiveness and accuracy of the proposed performance-evaluation metric for large-scale fiber tracking. Introduction In materials science and research, an important problem is to quickly and accurately reconstruct and characterize the underlying microstructure of a material sample [6]. For fiber-reinforced composite materials, the microstructure feature of interest is the fibers, whose shapes, orientations, and distributions directly affect the mechanical properties [10, 13]. One typical approach to reconstruct the large-scale fibers is to serial section the 3D material sample, take high-resolution microscopic images for each slice, and finally detect/track all the fibers through the slices [14, 16]. However, dense sampling of serial-sectioning, i.e., with very small inter-slice intervals, is time consuming and prevents the quick processing of large-sized material sample. On the other hand, overly sparse sampling of serial-sectioning, i.e., with very large inter-slice intervals, introduces uncertainty and ambiguity in tracking fibers across slices. One effective approach to address this issue is to dynamically adjust the sampling rate, i.e, the interslice intervals, in serial-sectioning by balancing the tracking accuracy with the data collection time. To achieve this goal, it is necessary to have a reliable evaluation of the tracking performance on the available slices with their inter-slice intervals before moving to serial-sectioning the next slice. This requires the use of an online tracking algorithm and the development of an on-the-fly tracking performance evaluation metric. By using an online tracking algorithm, such as Kalman filters, the fibers can be tracked based on the available slices and updated as soon as a new serial-sectioned slice is available. The on-the-fly performance evaluation continuously and quickly assesses the fiber-tracking performance as soon as a new slice is serial-sectioned and the online fiber tracking is updated using this new slice. The estimated tracking performance can then be used to adjust the inter-slice interval for the next slice. In material science, materials are reinforced by embedded objects such as small particles, fibers, or boundaries between different crystals. In order to be effective, these objects must be fairly dense, leading to a common characteristic that the microscopic structure is composed of crowded or densely packed mixtures of embedded objects. The objects are mainly fibers in this paper. The fact that the fibers are crowded induces a spatial consistency for fiber neighbors in different slices, so we take advantage of that spatial consistency in estimating the performance of large-scale fiber tracking. In this paper, we focus on addressing the problem of on-the-fly performance evaluation such that the fiber tracking performance can be quickly estimated without any interruption after a new slice is serial sectioned and the online tracking is extended to this new slice. In particular, this requires to exclude the human interactions from the tracking performance evaluation. Unfortunately, existing multi-target tracking performance evaluations metrics, such as the widely used Multiple Object Tracking Accuracy (MOTA) [5], usually require the ground-truth tracking results. For fiber tracking, these evaluation metrics count the coincidence between the tracked fibers and ground-truth fibers to evaluate the tracking performance. The construction of the ground-truth fibers requires manual annotation of fibers on each slice and manual linking of the fibers between slices. Manually annotating ground truth for large-scale fibers (about 500 similar-appearance fibers in one slice) in an image sequence makes the whole system not automatic. The objective of this paper is to develop a new metric that can evaluate the performance of online large-scale fiber tracking without requiring the ground truth. Related Works In most of the previous work [15, 9, 1, 8, 14], performance evaluation for object tracking requires manually annotated ground-truth tracking. By comparing generated tracking results with the ground-truth tracking on each image through the 142 IS&T International Symposium on Electronic Imaging 2017 Computational Imaging XV https://doi.org/10.2352/ISSN.2470-1173.2017.17.COIMG-437 © 2017, Society for Imaging Science and Technology