Skip to content

Data model

Field-level reference for the media_classification app's core models — the Classification four-type hierarchy, its dynamic attributes, and the AI provider models. For the concepts behind these, see Classification model and AI pipeline architecture.

Classification

Field Type Notes
resource FK → Resource The media item this classification is about.
classification_type choice FINAL | AI | USER | FEEDBACK.
model FK → AIProvider, nullable Set only for AI-type rows — which provider produced it.
source_classification FK → self, nullable For USER: the row it was forked from. For FINAL: the approved AI/USER row. For FEEDBACK: the approved row it corrects.
final_classification FK → self, nullable For AI/USER/FEEDBACK: always points at the resource's FINAL row.
frames_msgpack FileField Per-classification zstd-compressed msgpack of per-frame object data — see Msgpack contract.
is_approved bool Only meaningful on FINAL.
is_blank bool No animal/object of interest present.
is_setup bool Camera-setup/test shot, excluded from most analytics.
is_gold_standard bool Marked as fully correct (objects identified, classified, and tracked) — eligible for benchmarking/training data.

ClassificationDynamicAttrs

One row per observed object (or expert-authored group) on a non-FINAL classification.

Field Type Notes
classification FK → Classification (related_name="dynamic_attrs")
observation_type choice human | animal | vehicle | blank | unknown | unclassified.
observation_level choice object (default, one row per tracked individual) | group (expert-authored aggregate, no bbox/individual_id, count = group size).
species FK → Species, nullable See Taxonomy & GBIF.
sex choice unknown | female | male.
age choice unknown | adult | subadult | juvenile | offspring.
behaviour choice undefined | grazing | browsing | rooting | vigilance | running | walking.
count int Default 1. For group rows, the expert-authored group size.
count_new int, nullable
individual_id string, nullable Human-assigned individual mark/tag. Must be null on group rows.
classification_confidence decimal (0–1), nullable
bboxes JSON list [xmin, ymin, width, height], relative coordinates, for the object's first occurrence. Must be empty/null on group rows.
first_frame_index int 0 for images; original-video-FPS frame index for video.
distance float, nullable Metres from camera at first occurrence — see Distance estimation. For video at ≥5 FPS, averaged over the first 3 frames.
attrs JSON dict Custom attributes defined by the project's Classificator.
attr_confidence JSON dict Per-attribute AI confidence scores.

Database constraints enforce the object/group split: group rows can't carry bboxes, can't have an individual_id, and must have count >= 1.

AI provider models

AIProvider is a django-polymorphic base with two concrete subclasses:

Model Distinguishing field Created by
TrapperAIProvider connection → AIProviderConnection sync_ai_models (bootstrapped from an AI Manager)
ExternalAIProvider none — bare AIProvider manually, for a model not registered on any connected AI Manager

Shared AIProvider fields:

Field Type Notes
name, version, description string
provider_type choice detection | classification | depth | pose.
video_support bool
object_based bool
crop_image bool Crop before classification — species-classification providers only.
remote_id UUID, nullable PredictionModel.id on the AI Manager.
categories JSON Camtrap DP-shaped attribute mapping: {attributeMappings: [{attributeType, labels: [{aiLabel, value, gbifSpeciesKey, speciesId, ...}]}]}. speciesId is computed by sync_ai_models — never hand-edit it.
keypoint_topology JSON Pose providers only: keypoint names/skeleton/default OKS anchors, stamped at sync time.
is_active bool Managed by sync; manually created providers stay active.
model_file_hash string (SHA-256), nullable Re-matches a provider across re-registrations with a new remote_id.
keep_synced bool Disable to preserve manual edits (e.g. hand-tuned categories) across future syncs.
last_synced datetime, nullable
token string Used in AI-callback URLs.

AIProviderConnection (one per AI Manager instance):

Field Notes
api_url Base URL of the AI Manager.
api_auth_login / api_auth_passw Credentials of the AI Worker user on that AI Manager.
trapper_instance_url This Expert's base URL, as seen from the AI Manager (callback target).
is_active

AIClassificationJob

One row per AI run submitted for a (project × collection) pair.

Field Notes
id UUID.
user_remote_task OneToOne → UserRemoteTask (tracks the remote job's lifecycle on the AI Manager).
ai_provider FK → AIProvider.
collection FK → ClassificationProjectCollection, nullable.
stage Internal remote-task stage name.
job_config JSON snapshot of exactly what was submitted — minimum_confidence, skip_empty, skip_missing_labels, tracker/depth config. The single source of truth for "what did this run actually do?"

See also