In one line
Built an AI pipeline that sorts 600,000+ camera-trap photos into "animal / no animal" automatically, replacing an estimated 300+ hours of manual sorting with an unattended run, so a friend's Bachelor thesis can count wildlife instead of scrolling through empty frames.
Context & problem
My friend runs seven camera traps for his Bachelor thesis (submission January 2027). They had produced 600k+ photos and keep growing monthly – most frames are empty (wind, vegetation), and sorting them by hand would eat the thesis. On top: old cameras with irregular burst counts and clocks that drift or reset after battery changes.
What I did
- Scoped it to what unblocks the thesis first: animal / no animal (species and birds as a later phase)
- Chose MegaDetector v5a (the most cited version, which matters for the methods chapter) and set it up on an existing RTX 2070 SUPER at my friend's parents' PC – no new hardware; remote access via Tailscale, OpenSSH and HelpWire
- Designed calibration: a ~1,050-photo labeled sample, stratified by location and day/night, to pick the confidence threshold on evidence (target: miss under ~2 % of animals) and state a false-negative rate
- Designed the EXIF scan: burst sequences grouped by time gap, not file count; clock drift corrected per camera with anchor points and linear interpolation; everything in UTC (DST switch handled)
- Designed the outputs for a non-technical user: a detections spreadsheet, a ~2 GB crops folder for fast review on any laptop, and a positives tree sorted by confidence; monthly incremental re-runs
- Lined up the data model with Camtrap DP 1.0, the open standard for camera-trap data, so the dataset can be published to GBIF
Result / impact
The pipeline sorts 600k+ photos into animal and no animal and replaces an estimated ~300 hours of manual sorting. Only the positives and a sample of the discards need human review.
Details
- My role: designed and built the pipeline; my friend does the fieldwork, labelling and review
- Compliance & data protection: photos of people are detected and excluded from the dataset; permits and landowner consent in place; backup before any processing
- Standards & methods: MegaDetector v5a, stratified ground-truth sample, Camtrap DP 1.0 (TDWG)