AI Wildlife Cameras: Put New Locations in the Test

Imagine a fox passing the same fence on three successive nights. The photographs contain an animal, but also a recurring background. When an AI model correctly labels another picture from that camera, the result leaves an important question open: how well would it recognise wildlife beside a different fence?

For a nature project, an ecotourism business or a creator explaining conservation technology, that question deserves space beside the impressive demonstration. Camera-trap AI should be evaluated against the setting where its results will be used. A test made from familiar cameras can answer a different question from a test made at unfamiliar sites.

First, identify the model’s actual job

Microsoft’s MegaDetector documentation distinguishes detection from species classification. MegaDetector finds objects in camera-trap images and assigns the broad categories animal, person or vehicle. It does not identify an animal’s species. A species classifier is a further component, rather than a capability supplied by the animal-detection box itself.

This distinction helps when evaluating a demonstration. An accurately placed box around a fox establishes that the detector located an animal. The fox label requires its own assessment. Ask which component produced each output and what was checked, particularly when a service combines several models behind one interface.

What an unfamiliar camera revealed in 2018

In the ECCV 2018 paper Recognition in Terra Incognita, Sara Beery, Grant Van Horn and Pietro Perona studied images from 20 camera locations. Their dataset allowed comparison between test photographs from locations represented during training and photographs from locations absent from training.

In one full-image classification experiment, which excluded empty images and did not use image-sequence information, top-choice classification error rose from 19.06% at familiar locations to 41.04% at unfamiliar ones. Cropping images around animals improved classification, but a location gap remained. These are results for the paper’s particular models and dataset, not current accuracy figures for every wildlife AI system.

The editorial lesson is about test design. A score needs a description of what was unfamiliar in the test. New photographs of an established camera view and photographs from a new camera location are different challenges.

Modern species tools still need a clearly described scope

Google’s SpeciesNet documentation, checked on 25 October 2026, describes an ensemble combining detection and classification. Its classifier uses a fixed collection of labels, including species, broader taxonomic groups and non-animal categories. The documented decision process can return a broader group when the evidence does not support a species-level output.

The documentation also describes geographical filtering when location information is supplied. That makes location settings part of the result’s history. Record which country or region was provided, and preserve a broader output as broader. Rewriting a family-level label into a particular species for an attractive caption adds specificity the system did not supply.

Build a test that resembles the proposed deployment

Consider a hypothetical woodland visitor centre planning an illustrated wildlife diary. It already has labelled photographs from three camera positions and wants to add a fourth. Its immediate decision is whether the proposed workflow can produce reliable species suggestions at that new position.

One useful evaluation would keep photographs from the fourth location out of any local training or tuning, then compare its results with independently reviewed labels. If the centre adjusts the workflow after seeing those answers, it should reserve further unseen material for the next evaluation. Google’s guidance on dividing datasets explains why repeatedly using test results to guide changes can make the evaluation less trustworthy.

For a pretrained model, distinguish “withheld from our local development” from “never present in the provider’s training data”. If the latter is unknown, say so. Success at one held-out camera would support that tested setting; it would leave performance at other untested sites to be established.

Keep related frames from a short capture sequence together when constructing the evaluation. Otherwise, near-neighbour views of the same encounter can appear on both sides of a local training–test split. This applies the grouped-data principle in scikit-learn’s evaluation guidance to the proposed camera project; it is not a claimed result for the fictional centre.

A practical test plan should state:

• The intended task: locating animals, suggesting species, or preparing records for expert review.
• The separation: which cameras, capture sequences and dates were excluded from local development.
• The reference: who checked the labels, how disagreements were handled and which images remained uncertain.
• The coverage: the lighting, animal sizes, partial views and species represented in the reviewed material.

Inspect what the filter would hide

A tidy gallery of successful detections leaves another part of the workflow unexamined: the images sent to the empty pile. MegaDetector’s user guide explicitly recommends testing on the user’s own data and recording the model version and confidence threshold when analysing false positives and false negatives.

For the visitor centre, include a separately reviewed sample of images the proposed filter would set aside. An overlooked animal cannot receive a correct species label downstream if its image never reaches classification or review. Keep processing failures separate from images judged empty; an unreadable file supplies no visual evidence either way.

Choose review rules around the purpose. A rare-species announcement deserves closer checking than an internal queue of possible animals. The centre could require an experienced reviewer to approve every public species claim while using AI to organise the material. That is an operating choice to evaluate, not a promised accuracy level.

Publish the observation with its limits intact

Keep the original photograph, camera identifier, capture time, model version, relevant settings, proposed label and reviewer’s decision connected. MaryChuks.com’s guide to making a citizen-science record useful explains why an identification needs supporting evidence and an understandable observation record.

For readers, a concise description of the evaluation can be more revealing than a headline percentage: which locations were new, which animals were represented, and whether discarded images were inspected. AI wildlife tools become easier to assess when the published story shows where the label came from and how it was checked.

Sources

• Beery, Van Horn and Perona: Recognition in Terra Incognita — ECCV 2018, including Table 1.
• Microsoft: camera-trap detection and classification and MegaDetector user guide.
• Google: SpeciesNet documentation and ensemble decision-making.
• Google: dividing datasets; scikit-learn: evaluation with grouped data.

Documentation checked on 25 October 2026. The visitor centre and its evaluation plan are hypothetical.