Benjamin Bustos’s research while affiliated with University of Chile and other places

We present a study and comparison of computer vision methods for the task of finding repetitive motifs in ancient Peruvian pottery. Under this context, the main difficulties for solving the task are that the motifs are in most cases highly repetitive, that the motifs corresponding to the same pattern are slightly different due to be hand-drawn, and that the amount of data available for training and testing purposes is scarce. We evaluate and compare several techniques: Template Matching, Segment Anything Model, Mask R-CNN, Faster R-CNN, RetinaNet, and YoloV8-s. We conclude that YoloV8-s and Retina-Net are the most effective techniques for the task, but the effectiveness for zero-shot detection is low for all evaluated techniques.

We extend the concept of worst-case optimal equijoins in graph databases to the case where some nodes are required to be within the k-nearest neighbors (kNN) of others under some similarity function. We model the problem by superimposing the database graph with the kNN graph and show that a variant of Leapfrog TrieJoin (LTJ) implemented over a compact data structure called the Ring can be seamlessly extended to integrate similarity clauses with the equijoins in the LTJ query process, retaining worst-case optimality in many relevant cases. Our experiments on a benchmark that combines Wikidata and IMGpedia show that our enhanced LTJ algorithm outperforms by a considerable margin a baseline that first applies classic LTJ and then completes the query by applying the similarity predicates. The difference is more pronounced on queries where the similarity clauses are more densely connected to the query, becoming of an order of magnitude in some cases.

The SPARQL standard provides operators to retrieve exact matches on data, such as graph patterns, filters and grouping. This work proposes and evaluates two new algebraic operators for SPARQL 1.1 that return similarity-based results instead of exact results. First, a similarity join operator is presented, which brings together similar mappings from two sets of solution mappings. Second, a clustering solution modifier is introduced, which instead of grouping solution mappings according to exact values, brings them together by using similarity criteria. For both cases, a variety of algorithms are proposed and analysed, and use-case queries that showcase the relevance and usefulness of the novel operators are presented. For similarity joins, experimental results are provided by comparing different physical operators over a set of real world queries, as well as comparing our implementation to the closest work found in the literature, DBSimJoin, a PostgreSQL extension that supports similarity joins. For clustering, synthetic queries are designed in order to measure the performance of the different algorithms implemented.

The analysis and understanding of artefact properties and their relationships is a key goal in archaeological analysis of cultural heritage objects. There are many aspects of concern, including shape properties of the objects, but also appearance properties stemming from paintings and ornamentations on the object surfaces. To date, these are considered by experts mostly holistically and on a per-object basis. We present an approach for the interactive visual exploration and correlation of shape- and ornament-based properties of a large collection of ancient vessels. Our approach allows to group objects by properties, and to relate them in side-by-side and bipartite graph displays. To this end, we define an encompassing set of feature descriptors, which are leveraged to cluster the objects by properties, selected by the user. Case studies show that a comparative overview of all objects effectively supports the discovery of interesting co-occurrences of shape and ornament properties. This way, our tool opens new possibilities for the domain analysis of cultural heritage object collections by data-driven visual exploration.

During the last years, many advances have been made in tasks like 3D model retrieval, 3D model classification, and 3D model segmentation. The typical 3D representations such as point clouds, voxels, and polygon meshes are mostly suitable for rendering purposes, while their use for cognitive processes (retrieval, classification, segmentation) is limited due to their high redundancy and complexity. We propose a deep learning architecture to handle 3D models represented as sets of image views as input. Our proposed architecture combines other standard architectures, like Convolutional Neural Networks and autoencoders, for computing 3D model embeddings using sets of image views extracted from the 3D models, avoiding the common view pooling layer approach used in these cases. Our goal is to represent a 3D model as a vector with enough information so it can substitute the 3D model for high-level tasks. Since this vector is a learned representation which tries to capture the relevant information of a 3D model, we show that the embedding representation conveys semantic information that helps to deal with the similarity assessment of 3D objects. We compare our proposed embedding technique with state-of-the-art techniques for 3D Model Retrieval using the ShapeNet and ModelNet datasets. We show that the embeddings obtained with our proposed architecture allow us to obtain a high effectiveness score in both normalized and perturbed versions of the ShapeNet dataset while improving the training and inference times compared to the standard state-of-the-art techniques.