1. Overview
wildtag.ai processes folders of camera trap images automatically - detecting animals, classifying species, and organising results for review. It runs on your computer. wildtag ships without a model built in: you download the one you want from the Models tab the first time you use it, while connected to the internet. After a model is downloaded it is stored locally and works fully offline. DeepFaune is the recommended model for UK and European wildlife.
The workflow has four stages:
2. Getting Started
wildtag_Setup.exe from the link provided (about 670 MB, since it bundles the Python runtime; models are downloaded separately later).wildtag_Setup.exe and follow the wizard (no administrator rights needed). It creates a desktop and Start Menu shortcut. Launch wildtag from the shortcut; it opens directly with no console window. If Windows shows a security warning, click "More info" then "Run anyway" - this is normal for software that is not yet code-signed.validation\ subfolder.
3. Models
wildtag does not come with a species model built in. The Models tab is where you download the model you want, once, while connected to the internet. After a model is downloaded it is stored on your computer and used offline from then on. Downloading a model is the first thing to do after installing wildtag.
What the Models tab shows
Each model is listed as a card giving its region, size, licence, and a short description, along with its current status:
| Status | Meaning |
|---|---|
| Not installed | Available to download. Click Download to fetch it. |
| Installed | Already on this computer and ready to use offline. A Remove button frees the disk space if you no longer need it. |
| Planned | In development, not yet available. Shown for information only; it cannot be downloaded and does not appear in the Run tab. |
Downloading a model
Click Download on the model you want. wildtag confirms the size, then downloads it with progress shown on the card. DeepFaune is about 1.1 GB and SpeciesNet about 500 MB, so allow a few minutes on a normal connection. When it finishes, the status changes to Installed.
Which model to choose
DeepFaune v1.4 is fast on any computer, works offline, and is well-calibrated for UK and European species - a good default for European surveys. SpeciesNet covers 2000+ species worldwide; choose it for global coverage or taxonomy-level results. The full global model can be slow per image, especially without a GPU, so setting a Geographic filter (your country) is recommended to speed it up. See Appendix A for details of each model.
4. Running the Pipeline
Select your image folder
Click Browse and navigate to your camera trap image folder. wildtag counts all JPEG images found inside, including subfolders.
Choose a model
Select a species model from the dropdown. Each model brings its own animal detector, so there is no separate detector to choose. Only models you have downloaded on the Models tab appear here; if the dropdown is empty, download a model first.
| Model | Best for |
|---|---|
| DeepFaune v1.4 | UK and European surveys (recommended); fast on any computer |
| SpeciesNet Global | Worldwide coverage or taxonomy-level rollup; use a geographic filter to keep it fast (a GPU helps for large projects) |
Set confidence thresholds
- Detection threshold - images where no detection exceeds this value are recorded as
empty. Default: 0.1. - Species confidence - detections below this threshold are labelled
low_confidence. Default: 0.6. Lower this to see more predictions; raise it for stricter results.
Geographic filter
Select a region to filter out implausible species predictions. For UK surveys, select GBR. This removes predictions like wolverine or raccoon and replaces them with the next most plausible UK species. Select None to use raw model predictions.
Performance options
- Processing device - CPU is always available. NVIDIA GPU (CUDA) appears automatically if detected and dramatically speeds up classification.
- CPU threads - defaults to one less than your total CPU count, leaving one core free for the rest of your computer.
- Checkpoint frequency - how often wildtag saves its progress during a run. If a long run is interrupted, it resumes from the last checkpoint rather than starting over. The default suits most projects; there is rarely a need to change it.
Output options
- Generate IDs - adds unique
image_idanddetection_idcolumns to the CSV. - Prepare validation folder - creates downscaled annotated images organised by species, ready for the Validate tab.
- Validation image quality - sets how large the downscaled validation images are. Higher quality makes review easier but produces larger files and bigger volunteer packages; lower quality is faster to make and send. The default is a good balance for most surveys.
Running
Click Run wildtag. The log panel shows live progress. With DeepFaune, a few dozen images take a couple of minutes on CPU. SpeciesNet does more work per image (a full detector, classifier, and geolocation pipeline over 2000+ species), so it is slower; setting a Geographic filter (your country) narrows the candidates and speeds it up considerably, and a GPU helps further for large projects.
5. Validating Results
Validation is where you review AI predictions and correct any errors. wildtag presents images as a gallery, one species folder at a time.
Opening a validation folder
Click Browse and select your project folder or the validation\ folder inside it - wildtag finds it automatically. The species dropdown shows only folders with images still needing review.
Reviewing images
Images appear in a grid with the AI confidence score below each one and a bounding box drawn around the detected animal. Click any image to open the correction dialog.
- If the label is correct, click Correct
- If wrong, select the right species from the dropdown and click Apply correction
To correct several images at once, hold Ctrl and click each one; selected tiles are highlighted in blue. Then click any selected tile to apply a single correction to them all together. This is much faster when a whole batch shares the same error.
The correction dropdown is ordered: unidentifiable first, then species alphabetically, then broad groups such as mustelid or lagomorph, then special labels.
If the species you need is not listed, use the "Not listed? Add a label" box in the correction dialog to add it. New labels are saved to validation/custom_species.txt in the project, stay available for the rest of the project, and are included in volunteer packages so everyone sees the same options.
Completing a batch
Once you have reviewed all images in a folder, click Mark batch complete. This writes your corrections to the validation CSV. The folder will not appear again in the dropdown once fully validated.
Repairing a validation folder
If any tiles show "Image not found", the folder's image list has got out of step with the images actually present (this can happen with folders made by older versions of wildtag). wildtag repairs this automatically whenever a validation folder is opened, and a Repair folders button on the Validate tab does the same on demand. The repair is safe: it never deletes images or loses validation work, and it keeps a backup of the original list.
How validation updates the master CSV
As you validate, wildtag writes corrections to validation.csv files inside each species folder. When you distribute or export, these corrections are merged into the master results_with_ids.csv, populating the correct_label and human_verified columns. Your master CSV always reflects your latest validation work.
6. Distributing and Collecting
The Distribute tab handles packaging validation folders to send to volunteer validators, collecting their corrections when they are returned, and exporting your favourite (hearted) images by species.
Part A - Prepare packages for validators
Select your validation folder and click Prepare zip packages. wildtag creates one zip file per species folder containing the annotated images, the validation spreadsheet, the wildtag app, and instructions. Each zip goes into a distribute\ folder next to your validation folder.
Send the relevant zip to each volunteer. You might assign one species to each person, or send multiple zips to the same person for different species.
What the volunteer does
Run wildtag.bat..._validation.csv from the species subfolder (the filename does not matter)Part B - Collect returned validated zips
When validators return their work, drop the files into the collect\ folder inside your project folder. A returned file can be a zip of the whole folder, or just a validator's ..._validation.csv under any name; wildtag merges by each detection's unique ID, so the filename and folder structure do not matter, and a file may contain any subset of rows. Then click Merge all returned validations. wildtag will:
- Show a preview: how many validated detections and corrections are coming in
- Note any rows that do not match this project (for example a file returned to the wrong project), which are ignored
- Ask you to confirm before merging
- Update your master validation CSVs and
results_with_ids.csv - Move processed files to
collect\processed\; unreadable files are left incollect\to retry
collect\ folder if it does not exist and tells you where to drop the zips.Part C - Export as Camtrap DP
Click Export Camtrap DP to write your validated results as a Camera Trap Data Package (Camtrap DP), the community standard for camera-trap data. wildtag creates a camtrapdp\ folder containing deployments.csv, media.csv, observations.csv and a datapackage.json descriptor. Because it follows the standard, this package can be shared, archived, or uploaded to GBIF and other standards-based tools without any further conversion.
deployment.csv file in your project folder, with one row per camera site giving its name, latitude, longitude, and deployment start and retrieval dates. Use the Save deployment.csv template button to create a ready-to-fill template, then complete it before exporting. Without it, export still works but the deployment details and map locations will be missing.Part D - Export favourite images
While validating you can mark standout images as favourites (the heart on each tile). The Distribute tab can then export these into a favourites\ folder, organised into one subfolder per species, containing only the images you hearted. This is handy for pulling out the best shots of each species for reports or presentations.
7. Summary
The Summary tab gives an at-a-glance overview of a project: what was processed, how validation is progressing, where files were written, and a breakdown of detections by species. It fills in automatically when you open it.
What was processed
Run statistics for the project: how many images and detections were processed, how many frames were empty, and which model was used.
Validation progress
How many detections have been validated so far and how many remain, so you can see how much review is left across the whole project.
How many detections per species
A table of detection counts per species. After validation this reflects your corrected labels, giving a quick species summary for the survey.
Confusion matrix
Once you have validated some images, this compares the AI's original predictions against your corrections, showing where the model agreed with you and where it did not. It is a quick way to see which species the model handles well and which it tends to confuse.
Where your files have been saved
The locations of the key outputs for the project (the results CSV, the validation folder, and any exports), so you can find them without hunting through folders.
Viewing detections on a map
If your images carry GPS metadata, or you have supplied deployment coordinates, click Launch map to open an interactive map of your detection locations in your web browser. A dropdown on the map lets you colour sites by species.
8. Understanding the CSV Output
wildtag produces a CSV with one row per detection - not one row per image. An image with two animals produces two rows; an empty image produces one row with NA for species fields.
Core columns
| Column | Description |
|---|---|
image_id | Unique identifier for the source image. Consistent across runs on the same image. |
detection_id | Unique identifier for this specific detection. |
relative_path | Path to the image relative to the survey folder root. |
label | Final species label: empty, human, vehicle, low_confidence, or a species name. |
confidence | Classifier confidence score (0-1). NA for empty images. |
correct_label | Correction entered during validation. Empty if the AI label was accepted. |
human_verified | TRUE once a detection has been reviewed in the Validate tab. |
Detection detail columns
| Column | Description |
|---|---|
detector_label | What the detector found: animal, human, or vehicle. |
detector_confidence | Detector confidence score. |
cv_label | Raw species label from the classifier before geofencing or thresholding. |
cv_confidence | Classifier confidence for cv_label. |
cv_model | Name of the classifier used, e.g. deepfaune-v1.4. |
bbox_left, bbox_top, bbox_right, bbox_bottom | Bounding box coordinates. Normalised (0-1) when bbox_normalised = 1. |
EXIF columns
wildtag automatically extracts camera metadata including DateTimeOriginal, Make, Model, GPS coordinates, ExposureTime, and ISOSpeedRatings.
correct_label where populated, falling back to label where not. This gives the best available identification for every detection. Note that results_with_ids.csv is produced by the Export results button on the Summary page (and automatically when you close wildtag); see Appendix C for how validation data is stored and exported.9. Privacy and GDPR
Camera traps occasionally capture people. wildtag handles this in two ways:
- Detection labelling - MegaDetector identifies humans and vehicles separately from animals, labelled
humanorvehiclein the CSV. - Automatic pixelation - validation copies of human detections are heavily pixelated. Faces and bodies are unrecognisable. Original images are never modified.
Appendix A: AI Models
MegaDetector v5a
Developed by Microsoft AI for Earth. A YOLOv5-based detector trained on millions of camera trap images worldwide. Identifies three categories: animal, human, and vehicle, with bounding boxes. Does not identify species.
- Architecture: YOLOv5x, 140M parameters, 1280px input
- Weight file: 280 MB
DeepFaune v1.4
Developed by CNRS (France). A vision transformer classifier trained on European camera trap images. Classifies animal crops into 38 classes covering the full range of European wildlife. Downloaded from the Models tab on first use (about 1.1 GB) and used offline afterwards.
- Architecture: ViT-Large with DINOv2 backbone, 182px crop
- Classes: 38 including domestic animals and birds as a group
- Weight file: 1.1 GB (downloaded on first use, not bundled)
- Best for: European surveys, ungulates and carnivores
SpeciesNet Global (Google, v4.0.3a)
Developed by Google. A complete end-to-end pipeline combining its own internal MegaDetector with a global species classifier and geofencing. Outputs taxonomy-aware predictions that roll up to family or order level when species-level confidence is insufficient.
- Architecture: EfficientNetV2-M classifier with internal MegaDetector
- Classifier weight file: 224 MB (downloaded on demand on first use, not bundled)
- Best for: global surveys or taxonomy-level rollup, on a machine with a GPU (very slow on CPU)
Appendix B: Geographic Filtering
The geographic filter reduces false positives by removing species predictions that are implausible for your study region. When GBR is selected, wildtag checks each prediction against a curated list of UK species. If the top prediction is not on the list, wildtag uses the next most confident plausible prediction instead.
UK species list includes
- Deer: red deer, roe deer, fallow deer, sika deer, Chinese water deer, muntjac
- Carnivores: red fox, badger, otter, stoat, weasel, polecat, pine marten, mink, wildcat
- Rodents: red squirrel, grey squirrel, water vole, bank vole, field vole, wood mouse, dormouse, brown rat
- Lagomorphs: rabbit, brown hare, mountain hare
- Other: hedgehog, mole, wild boar, all shrews, domestic livestock, dog, cat, bird
Broad group labels such as mustelid, lagomorph, and bird always pass the geofence since they are plausible anywhere. Select None to disable geofencing entirely.
Appendix C: How wildtag Stores and Updates Validations
This appendix explains what happens to your validation work behind the scenes. You do not need to know any of this to use wildtag, but understanding it gives you confidence that your corrections are saved safely, that the results file is always reproducible, and that nothing is lost if the program closes unexpectedly.
Two kinds of data, kept apart
A camera-trap project holds two very different kinds of information. The first is the detection record: every animal wildtag found, with its image, location, time, box, and predicted label. For a large survey this can run to hundreds of thousands of rows, and once a run has finished it does not change. The second is your validation decisions: the handful of labels you confirm or correct as you review images. This is small, but it changes constantly as you work.
wildtag keeps these two things in separate places. The detection record stays in results_with_ids.csv, written once when you run the pipeline. Your validation decisions go into a small validation store inside your project folder, updated the instant you make them. Keeping the frequently-changing part small and separate from the large, fixed part is what makes validating feel immediate even on very large projects.
validation.db, kept in your project folder next to results_with_ids.csv. It is part of your project: keep it with the project and include it in your backups.What happens while you validate
Each time you mark a batch complete, or collect returned volunteer files, wildtag records those decisions in the validation store straight away. Because the store is small, this is effectively instant, there is no waiting while a large file is rewritten. The Summary page and the confusion matrix read directly from the store, so they always reflect your most recent work the moment you open them.
The results file is produced on demand
The combined results_with_ids.csv you use for analysis, the one where corrected labels appear in the correct_label column, is rebuilt from the detection record plus your validation store whenever you ask for it. You do this with the Export results button at the top of the Summary page. Exporting is the only moment wildtag rewrites the large file, so it is a deliberate step you control rather than something that slows down every correction.
results_with_ids.csv is regenerated from your validation store, avoid hand-editing it in Excel: your edits would be overwritten the next time you export. The validation store is the source of truth. To validate, use the Validate tab; to get a fresh spreadsheet, click Export.Checks when you open and close a project
wildtag looks after keeping your results file in step with your validations at two moments:
Why this matters
- Speed. Validating never waits on the large file, so it stays fast however big your survey grows.
- Safety. Your validation decisions live in their own store and are written the moment you make them, so an unexpected close does not lose recent work.
- Reproducibility. The results file can always be rebuilt from the detection record and the validation store, so it is never the only copy of anything important.