New World, New Worlds: Visual Analysis of Pre-columbian Pictorial Collections

Daniel Gatica-Perez, Edgar Roman-Rangel, Jean-Marc Odobez, Carlos PallanMM4CH 2011 · DOI

TL;DR

This paper introduces the CODICES project, which aims to develop computational tools for analyzing Maya hieroglyphic collections. They present a novel shape descriptor, HOOSC, and a new dataset of Maya syllabographs, demonstrating their effectiveness in retrieving and classifying these complex ancient symbols.

What Problem Does This Paper Address?

The study of ancient Maya civilization, particularly its intricate hieroglyphic writing system, has historically relied on manual analysis by experts. While digital technologies have enabled the creation of vast collections of Maya artifacts, monuments, and buildings, the sheer volume of this data poses a significant challenge for researchers. Manually cataloging, analyzing, and retrieving specific glyphs from these digital collections is time-consuming and prone to human error. This paper addresses the need for automated and semi-automated tools to manage, analyze, and search these rich visual archives, thereby accelerating research and improving accessibility to Maya cultural heritage.

Methods

The CODICES project focuses on visual analysis techniques for Maya hieroglyphs. Their core technical contribution is the HOOSC (Histogram of Orientation Shape-Context) descriptor. This descriptor combines two established computer vision approaches: Shape Context (SC) for capturing spatial relationships of points on a shape's contour, and Histograms of Oriented Gradients (HOG) for representing local orientation distributions. HOOSC represents a shape multiple times from different sampled points on its contour, aggregated into a robust description. For efficient comparison of glyphs, they employ a bag-of-visual-words (BoVW) representation, where quantized descriptors are clustered to form a vocabulary, and glyphs are compared based on the distribution of these visual words. They also discuss an improved version of HOOSC that incorporates contour thinning, efficient key point selection, and explicit spatial positioning for enhanced performance.

Datasets

The paper introduces a newly compiled Maya syllabic dataset. This dataset consists of 1270 segmented Maya syllabographs, categorized into 24 visual classes based on their Thompson catalog numbers. This is presented as the largest dataset of Maya syllabographs analyzed with automatic techniques to date. The dataset is divided into "candidate" (80%) and "query" (20%) subsets for evaluating the generalization capabilities of computer vision methods. The sources for this dataset include the AJIMAYA project, the Macri and Looper syllabic catalog, the Thompson catalog, and the FAMSI website.

Key Insights and Ideas

  • The complexity of Maya hieroglyphs extends beyond simple visual representation: The paper highlights that Maya glyphs, whether syllabographs or logographs, can be visually rich and complex. Phenomena like conflation, infixation, superimposition, and pars pro toto further increase the challenge of visual analysis.
  • Bridging the gap between cultural heritage and computer vision: The CODICES project demonstrates how advanced computer vision techniques can be applied to preserve, research, and educate about cultural heritage, specifically Maya hieroglyphics.
  • The importance of specialized descriptors for cultural artifacts: The HOOSC descriptor, tailored for the unique characteristics of Maya glyphs, shows superior performance compared to general-purpose shape descriptors.
  • The potential for automated tools to democratize access to ancient knowledge: By developing tools for visual retrieval and analysis, the project aims to empower a wider range of researchers and students to engage with Maya inscriptions.

Key Results

The paper reports promising results for the HOOSC descriptor:

  • In initial comparisons on a smaller dataset, HOOSC achieved a mean average precision (mAP) of 0.39, outperforming Shape Context (0.322) and Generalized Shape Context (0.279).
  • On the larger Maya syllabic dataset, an improved version of HOOSC achieved a MAP of 0.54.
  • Qualitative results show that the improved HOOSC descriptor successfully retrieves relevant glyphs within the top positions for query examples.

Limitations and Future Work

  • Focus on syllabographs: The current dataset and analysis primarily focus on syllabographs, with logographs reserved for future work due to their greater complexity and number.
  • Performance drop with larger datasets: The initial HOOSC descriptor experienced a drop in retrieval precision (almost 10%) when evaluated on the larger Maya syllabic dataset compared to a smaller one. This motivated the development of an improved version.
  • Future work includes:
    • Developing tools for automatic analysis of visual variability to study the evolution of inscriptions through time and across regions.
  • Creating a visual retrieval system that allows archaeologists to quickly search large collections using visual queries.
  • Expanding the analysis to include logographs and more complex inscription structures.
  • Further refining the HOOSC descriptor and exploring other shape representation approaches.

Summary updated

BibTeX
@inproceedings{gaticaperez2011new,
  author = {Daniel Gatica-Perez and Edgar Roman-Rangel and Jean-Marc Odobez and Carlos Pallan},
  title = {{New World, New Worlds: Visual Analysis of Pre-columbian Pictorial Collections}},
  booktitle = {MM4CH},
  year = {2011},
  doi = {10.1007/978-3-642-27978-2\_8},
}
Figure