Fine-grained Recognition Datasets for Biodiversity Analysis

Erik Rodner, Marcel Simon, Gunnar Brehm, Stephanie Pietsch, Johann-Wolfgang Wägele, Joachim DenzlerarXiv 2015 · arXiv

Abstract

Biodiversity research faces a significant challenge: the volume of species to be identified far outpaces the availability of human experts. This paper introduces two new datasets for fine-grained visual classification (FGVC) focused on tropical moths and butterflies, featuring up to 675 highly similar species. By applying Convolutional Neural Networks (CNNs) and spatial pyramids, the authors demonstrate that while current computer vision techniques show promise, the sheer scale and subtlety of biodiversity data present a significant frontier for the field.

Introduction: The Biodiversity Bottleneck

The identification of species in tropical hotspots is a labor-intensive process that currently acts as a bottleneck for ecological research. Many species remain unknown to science, and manual annotation requires specialized expertise. While the computer vision community has advanced fine-grained recognition—distinguishing between very similar categories like bird species—most existing benchmarks utilize a relatively small number of classes. This research bridges the gap between ecology and AI by providing datasets that reflect real-world biological complexity, testing whether standard deep learning tools can handle hundreds of nearly identical species.

Methodology

The authors established a baseline for these new datasets using a combination of deep learning and traditional machine learning techniques:

  1. Feature Extraction: Features were extracted from the pool5 layer of an AlexNet architecture pre-trained on ImageNet.
  2. Classification: These features were processed using a one-vs-all linear Support Vector Machine (SVM).
  3. Spatial Pyramids: To capture localized differences in wing patterns, a two-level spatial pyramid was implemented. This involved calculating global features for the entire image and local features from four equal subregions, which were then concatenated for final classification.

Datasets: Tropical Lepidoptera

The study introduces two primary datasets featuring specimens with "artificially spread wings," the standard scientific preparation for museum collections.

A specimen of Arsenura batesii showing the standard scientific preparation with spread wings.
  • Ecuador Moth Dataset: Comprises 2,120 images across 675 species from the Geometridae family. This set is particularly challenging due to numerous "look-alike" species, which were verified using DNA barcoding.
  • Costa Rica Dataset: Includes 3,224 images of 331 species of moths and butterflies, offering a higher image-per-category ratio than the Ecuador set.
  • Baseline Comparison: The Leeds Butterfly Dataset (10 categories) was used to demonstrate the relative ease of low-class-count problems compared to these new benchmarks.

Key Results and Insights

The performance of the baseline models highlights the increasing difficulty of classification as the number of species scales:

  • Ecuador Dataset (675 classes): Achieved 55.7% global accuracy. While high for the number of classes, it indicates substantial room for improvement.
  • Costa Rica Dataset (331 classes): Achieved 79.5% global accuracy, rising to 82.1% with the spatial pyramid approach.
  • The Complexity Gap: The same techniques achieved 99.24% on the 10-class Leeds Butterfly dataset, emphasizing the challenge of species-level identification at scale.
Examples of correct (green) and incorrect (red) classifications across various species, highlighting the difficulty of distinguishing look-alike species.

Several key insights emerged from the analysis:

  • The "Long Tail": Biological datasets often suffer from a scarcity of examples per species, essentially making the task a "few-shot learning" problem.
  • DNA Barcoding: Mitochondrial DNA analysis provided a ground-truth accuracy that surpasses expert visual inspection for cryptic species.
  • Local vs. Global Features: While spatial pyramids improved accuracy for the Costa Rica dataset, they slightly hindered performance for the Ecuador set, suggesting that global pattern arrangements may be more discriminative for certain moth families.

Future Directions

The authors identify several critical hurdles for future research:

  • Open-set Recognition: Models must be developed to identify previously unseen species rather than forcing them into existing categories.
  • Few-shot Learning: Systems that can learn from just one or two examples are vital for identifying rare taxa.
  • Interpretability: Biologists require models that provide "interpretable features" corresponding to actual biological traits, such as specific wing venation or spots.
  • Human-in-the-loop: Integrating active learning could allow AI to collaborate with experts on ambiguous cases, accelerating the cataloging of global biodiversity.

Summary updated

Abstract

In the following paper, we present and discuss challenging applications for fine-grained visual classification (FGVC): biodiversity and species analysis. We not only give details about two challenging new datasets suitable for computer vision research with up to 675 highly similar classes, but also present first results with localized features using convolutional neural networks (CNN). We conclude with a list of challenging new research directions in the area of visual classification for biodiversity research.
BibTeX
@misc{rodner2015fine,
  author = {Erik Rodner and Marcel Simon and Gunnar Brehm and Stephanie Pietsch and Johann-Wolfgang Wägele and Joachim Denzler},
  title = {{Fine-grained Recognition Datasets for Biodiversity Analysis}},
  year = {2015},
  eprint = {1507.00913},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1507.00913},
}
Figure