Insect Identification in the Wild: The AMI Dataset

TL;DR

This work provides the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists, and train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups.

Abstract

Insects represent half of all global biodiversity, yet many of the world's insects are disappearing, with severe implications for ecosystems and agriculture. Despite this crisis, data on insect diversity and abundance remain woefully inadequate, due to the scarcity of human experts and the lack of scalable tools for monitoring. Ecologists have started to adopt camera traps to record and study insects, and have proposed computer vision algorithms as an answer for scalable data processing. However, insect monitoring in the wild poses unique challenges that have not yet been addressed within computer vision, including the combination of long-tailed data, extremely similar classes, and significant distribution shifts. We provide the first large-scale machine learning benchmarks for fine-grained insect recognition, designed to match real-world tasks faced by ecologists. Our contributions include a curated dataset of images from citizen science platforms and museums, and an expert-annotated dataset drawn from automated camera traps across multiple continents, designed to test out-of-distribution generalization under field conditions. We train and evaluate a variety of baseline algorithms and introduce a combination of data augmentation techniques that enhance generalization across geographies and hardware setups.

Figures (8)

• Figure 1: Illustration of the difference between insect species identification from human-captured images and in the wild. On left are shown images of three moth species from our AMI-GBIF dataset, curated from citizen science data and museum collections. On right are shown images of the same species (along with other insects) within a photograph taken by an automated camera trap in the wild from our AMI-Traps dataset. One of the challenges of the AMI dataset is generalizing from AMI-GBIF to AMI-Traps without additional labeled data.
• Figure 2: Examples of challenges in the AMI-Traps dataset. (a) Incomplete: In some occasions, the body of an insect is incomplete due to its position in the automatically taken image. (b) Overlapping: Sometimes insects occlude each other. (c) Motion: Moving insects can be blurred and hard to classify. (d) Lighting: The lighting and camera exposure of the images is variable and can lead to poor contrast that masks some or all detail. (e) Resolution: Some insects are very small, leading to low-resolution images. (f) Perspective: While insects are generally positioned at a fixed distance, sometimes insects appear in the air or perched on the camera itself. Additionally, unexpected objects such as spider webs, dirt, and vegetation can be captured by the cameras and may be confused with objects of interest.
• Figure 3: Sample images from the AMI Dataset. The AMI Dataset is composed of (1) AMI-GBIF, curated from a number of sources with imagery from citizen science and museum collections, and (2) AMI-Traps, drawn from automated camera traps for insects across five countries in three regions. The number of individual insects annotated per sub-dataset is denoted in parentheses.
• Figure A.4: Examples of pictures removed during the dataset cleaning: (a) Contains multiple specimens referenced by multiple observations, (b) a placeholder for more than 100,000 observations, (c) an unsuitable picture containing only a moth body part, (d) a thumbnail, and (e) an animal in the larval life stage.
• Figure A.5: Distribution of geographic locations for images in AMI-GBIF. Since we consider species from the NE-America, C-America, and W-Europe regions, observations are concentrated in these regions. However, some observations fall elsewhere, reflecting the fact that some species present in a region of interest are also found elsewhere; all global observations of that species are drawn upon in creating AMI-GBIF.