Event-Based Insect Classification

This dataset comprises event-based recordings of individual insect flight trajectories collected under controlled outdoor conditions using a catch-and-release protocol. All insects were captured on the grounds of the University of Hohenheim (Stuttgart, Germany) in 2024 and 2025 using butterfly nets and released unharmed after recording.

Recordings were performed using a stereo event-camera setup positioned inside an open-sided structure at the State Institute for Bee Research Hohenheim. A uniform wall served as background to reduce non-insect events and shadows. The camera was mounted at approximately 65 cm height and 1.40 m distance to the wall; additional recordings at 1.80 m distance were included to increase variability. Each insect was recorded individually and released just outside the camera’s field of view to capture natural take-off trajectories. A barrier was introduced behind the release point to increase successful flight captures.

Raw event streams were manually segmented to extract individual insect flight trajectories. Temporal scaling was standardized, and background events were removed using 3D spatiotemporal cropping. The resulting foreground-only trajectories were restored to the original temporal resolution and stored as separate CSV files containing spatial coordinates (x, y), polarity (p), timestamp (t), and event index. Trajectories are named according to pollinator group, individual ID, stereo camera (left/right), and segment index.

The dataset includes 783 trajectories comprising approximately 82 million events with a total duration of 497 s. It covers six pollinator groups: honey bees, bumble bees, wild bees, hoverflies, butterflies, and other insects. Data are split into training (564), validation (112), and test (107) subsets with approximately balanced class distributions and no trajectory overlap between splits.

All files are organized into training/, validation/, and test/ directories and are free of background activity.