Turn a season of wildlife-camera photos into a searchable animal record

A wildlife camera can leave its owner with a familiar problem: plenty of photographs and no easy answer to which animals visited, when they appeared or how activity changed through a season. AddaxAI brings detection, species suggestions and review into one desktop application. It is free and open source, with Windows, macOS and Linux versions. Processing runs locally once the required software and models are available.[1]

For a garden group, reserve volunteer or small monitoring project, the useful outcome is a record that can be searched and corrected. The walkthrough below follows the current documentation; it is a guide to the available workflow, rather than a report of a field trial.

Start with one camera and a manageable folder

Begin with an untouched copy of the photographs and a small working folder containing a mixture of daylight, night-time, empty and animal images. This gives you a practical sample for checking whether the chosen model suits your camera position and local wildlife before committing the entire season.

In Analyse a folder, select the folder, a detector and a species classifier. Run the analysis, review the labels, then use Save to export the results. This simple route is useful for learning the interface or processing a single collection. Projects add the structure needed for repeated deployments and event counts.[2]

The two model choices do different jobs. AddaxAI currently recommends MegaDetector v5a as its default detector: it finds animals, people and vehicles in an image. A species classifier then examines the detected animal. The model zoo offers a global starting point in SpeciesNet and regional alternatives; its Europe-focused Deepfaune 1.5 entry, for example, lists 40 classes. Check the actual class list against the animals you expect. A model cannot supply a species label that its supported categories do not include.[3]

Give the season a structure

For a continuing survey, create a project and add sites. Within each site, make a deployment for each camera placement and recording period, pointing it to the relevant folder. A project keeps these deployments together, supports queued analysis and provides dashboards for activity, trends and maps. Site coordinates enable mapping. This structure preserves where a sighting came from when several cameras or seasons are combined.[4]

Use consistent names from the start: a site such as Orchard, then a deployment identified by camera and date range. Keep a simple record of when each camera was actually working. These small organisational choices make it easier to distinguish a quiet location from a camera that stopped recording.

Review species first, then the number of animals

In Labels, the Detections view works through individual boxes; Files shows the whole photograph or video. Correct a mistaken species, remove a false detection or draw a missing box. Also inspect a sample through Files with Show only empty selected, because reviewing only the predicted animal categories will miss anything the detector overlooked. Broad category review can help find animals assigned to the wrong species.[5]

These corrections survive changes to filtering and thresholds. Re-running AI is different: the documentation says it analyses the material from scratch and overwrites corrections. Finish the model choice on the small sample before investing heavily in review of the full project.[5]

Counts is a separate project step. An event groups files from one camera using the chosen independence interval. The automatic count for a species is the largest number visible together in a single image within that event. Open the event, inspect the other frames and confirm or amend the suggestion. Until confirmation, the AI suggestion remains the count.[6]

That distinction matters when a camera fires repeatedly. Twelve photographs of a fox walking past may describe one visit. They do not automatically describe twelve foxes. Equally, animals arriving one after another may require a human adjustment when the full sequence shows more individuals than any single frame. Treat events as encounters under your stated grouping rule; they are not a census of unique animals across the season.[6]

Make the first comparison small and interpretable

Check exported trap nights against the deployment log and remove known camera downtime before calculating encounter rates. The automatic field uses the inclusive span from first to last file, which does not independently confirm continuous camera operation.[7]

An illustrative calculation shows why camera effort belongs beside the count. Suppose Orchard records six fox events over 20 operating nights and Stream records nine over 45. Events divided by nights, multiplied by 100, gives 30 and 20 events per 100 nights respectively. Stream has more events in total; Orchard has the higher recorded encounter rate. Those numbers are invented to show the calculation, not measured results from AddaxAI.

Use the same event interval and a comparable camera protocol, then examine the photographs behind an apparent change. A useful first seasonal report can be modest: which species were recorded, the reviewed encounter rates and the times activity concentrated. Once that record is sound, the next camera download extends it instead of starting another unsorted archive.

Report an error or suggest a correction