Video Analysis of AI Detector Performance on Anomalous Subject

A 2013 video analysis evaluates AI object detection on a single anomalous subject. John Kosinski published this case study in 2026, containing 111 original frames, 111 cropped frames, annotated images, and structured detection tables. The data documents the MegaDetector and SpeciesNet system's response to an unidentified, approximately 8-foot-tall subject.

Use Cases

• Compare detector performance between full-frame and cropped-frame images for a single subject.
• Analyze SpeciesNet JSON outputs for 'abstain' classifications on subjects outside its training taxonomy.
• Evaluate bounding box annotations and structured detection tables for label alternation between 'animal' and 'human'.
• Correlate AI detection outcomes with on-site height verification data (approximately 8 feet).

Strengths

• Provides 111 matched pairs of full and cropped frames for controlled comparison.
• Includes multiple data modalities: images, JSON outputs, annotated images, and a structured table.
• Contains ground truth verification from a follow-up site visit confirming subject height.

Limitations

• Extremely small sample size focused on a single video (VID00002) and subject.
• Analysis is from a specific, singular event on 2013-07-29, limiting generalizability.
• Unknown number of rows in the structured detection table.

Provenance

Source

John Kosinski via figshare.

Collection Method

Frame-by-frame analysis of video using MegaDetector and SpeciesNet, supplemented by on-site verification.

Time Range

Primary video recorded on 2013-07-29; analysis published 2026-04-04.

Freshness

Analysis published in 2026 on data from 2013.