Glossary¶
Terms used across these docs that mean something specific in TRAPPER, alphabetically.
- AI Classification Job — One run of the AI pipeline submitted for a (Classification Project × Collection) pair. Tracked with a UUID, a
job_configsnapshot of exactly what was submitted, and a link to aUserRemoteTaskfollowing the remote job's lifecycle on the AI Manager. See Run & re-run the AI pipeline. - AI Provider — A configured AI model available to a Classification Project — either a
TrapperAIProvider(backed by an AI Manager connection) or anExternalAIProvider(manually configured, no connection). Has aprovider_type(detection/classification/depth/pose) and acategoriesJSON describing what it can predict. See Data model. - AI Provider Connection — The credentials + URL pair pointing at one AI Manager instance. Multiple
TrapperAIProviders can share one connection. See Configure an AI Manager connection. - Calibration sample — A
DeploymentCalibrationSample: one reference Resource on a Deployment with marked(bbox, distance, frame_index)points at known real-world distances. Fitting the calibration action over a Deployment's samples produces its Deployment calibration result. There's no separate "calibration deployment" type — any Deployment can have calibration samples attached directly. See Distance estimation. - Camtrap DP — The TDWG camera-trap data exchange standard. TRAPPER exports to it and aligns AI provider
categoriesJSON and Sequence/event grouping to its vocabulary. See explanation. - Classificator — The schema of attributes (species, sex, age, custom fields, …) that a Classification Project's annotation form is built from. One Classificator can be reused across multiple projects.
- Classification — The unified model for AI predictions, human edits, and post-approval feedback — see Classification model for the four types (FINAL/AI/USER/FEEDBACK).
- Classification Project — The unit that ties a Research Project, a Classificator, one or more Collections, AI Providers, and annotators together. See Create a classification project.
- Collection — A grouped set of Resources, typically one ingestion session (e.g. one SD-card's worth of footage). Becomes the on-disk top-level directory name when ingested via
trapper-tools. - Coordinator — The host-level process (
trapperai-coordinator, packagetrapperai-core) that runs on the AI Worker machine, not in a container — so it can access GPU/Hailo hardware directly. Dispatches jobs to the installed Runtimes. See AI pipeline architecture. - Deployment — A camera placement with a start/end time at a Location — one camera-out-to-pickup session.
- dynamic_attrs — Shorthand used throughout the code and these docs for
ClassificationDynamicAttrs— the per-object (or per-group) observation row attached to an AI/USER/FEEDBACK Classification. See Data model. - FEEDBACK (classification type) — A correction logged against an already-approved Classification. Doesn't reopen approval — it's a permanent QC trail.
- FINAL (classification type) — The container row representing a resource's approved-or-pending state within a project. Has no observation data of its own — reads it via
source_classification. - frames_msgpack — The zstd-compressed MessagePack file holding per-frame object data for one Classification. See Msgpack contract.
- Gold standard — A Classification (
is_gold_standard=True) reviewed as fully correct — every object identified, classified, and tracked. Used for benchmarking and training data, not a routine annotation state. - Group (observation level) — A
ClassificationDynamicAttrsrow an expert adds by hand to record an aggregate group size, with no bounding box and no individual track. Opposite of the defaultobjectlevel. - hypertable /
ObjectFrameObservation— A TimescaleDB-backed queryable index of per-frame data, built fromframes_msgpack. The extension itself is mandatory; what's on-demand is each project's snapshot being populated into it. Never the source of truth — see Frame timeseries. - Predictor — A model-execution unit inside an AI Worker Runtime package (e.g. a depth predictor, a species classifier). Distinct from a Tracker, which links detections across frames rather than producing them.
- Research Project — The top-level container for a body of camera-trap work — owns Locations, and the Collections/Classification Projects built on top of them.
- Resource — A single uploaded media file (image or video) within a Collection.
- Runtime — An installable AI Worker package providing model-execution code for specific hardware/frameworks (e.g.
trapperai-trackers-ultralytics,trapperai-predictors-torch-depth,trapperai-predictors-hailo-depth). Installed alongside the Coordinator based on detected hardware. - Sequence — An automatically-built grouping of Resources within a Collection into one ecological "event", based on a configurable time interval between consecutive timestamps. Mirrors Camtrap DP's
observations.eventID. - Tracker — Worker-side logic that links per-frame detections into a multi-frame track (a
DetectedObject). See Configure trackers — and that page's caveat about which tracker types are actually user-selectable. - USER (classification type) — A human-authored or human-edited Classification, often forked from an AI Classification via
source_classification.