01 · Image captioning
Detailed Image Captions
Annotators describe the whole scene in complete sentences: who or what is present, what they are doing, how they interact and what surrounds them. A generic label such as "car" or "street" becomes a description that a captioning or retrieval model can learn from. Caption length, tone and vocabulary follow your guidelines, whether you need one factual sentence or a fuller paragraph covering secondary detail.
Main subjectsActionsInteractionsSurroundingsContext
ExampleA red sedan is driving along a busy urban street while pedestrians cross near a commercial building.
02 · Scene classification
Scene and Environment Classification
Each image is classified by setting and conditions: indoor or outdoor, urban, rural or natural, industrial or commercial, plus weather, lighting and time of day. Categories can be single-label or multi-level (for example outdoor, natural, mountain), and we apply your taxonomy with decision rules for borderline images so that labels stay comparable across thousands of files.
Indoor / outdoorUrban / ruralWeatherLightingDay / night
ExampleOutdoor, mountain landscape, forest, lake, cloudy weather, daytime.
03 · Object attributes
Object Attributes and Characteristics
For every object that matters to your use case, annotators record visible characteristics: category, colour, shape, size, material, texture, condition and any task-specific attribute. The attribute list is yours to define. A retail catalogue needs different fields from an agricultural survey, so we configure the taxonomy to the project instead of forcing a fixed set.
CategoryColour & shapeSizeConditionCustom fields
ExampleBicycle – red, two wheels, metal frame, parked.
04 · Object relationships
Object Relationships
Relationship annotation records how objects and subjects relate: a person riding a bicycle, a person holding a tool, a car parked in front of a building, an animal standing near a tree. This is what separates a list of detected things from an understanding of what is happening in the image, and it is the layer that models trained only on plain object labels tend to lack.
ActionsInteractionsOwnership & contactProximity
ExamplePerson → riding → Bicycle
Car → parked in front of → Building
05 · Spatial relationships
Spatial and Relational Directions
Annotators record where objects sit relative to each other: left or right, above or below, in front of or behind, near or far, inside or outside, foreground or background, overlapping or adjacent. These are captured as structured labels from a controlled vocabulary, with a written rule for the reference point (the viewer's left or the subject's left) so that the same layout always gets the same label.
Left / rightAbove / belowFront / behindForeground / backgroundOverlapping
ExamplePerson: left of car. Tree: behind person. Vehicle: foreground.
06 · Scene graphs
Scene Graph Generation and Annotation
A scene graph organises an image into entities, their attributes and the relationships between them, so that it can be read by a machine and not only by a person. We build graph-style annotation as subject, predicate, object records tied to object IDs, ready to load into a scene-understanding, visual question answering or graph-based pipeline. A worked example appears in the next section.
EntitiesAttributesRelationshipsObject IDs
ExamplePerson → riding → Bicycle
Bicycle → beside → Car
Car → on → Road
07 · Materials & surfaces
Materials and Surface Identification
Visible materials and surface qualities are tagged per object or per region: wood, metal, glass, concrete, fabric, plastic, polished or unpolished, smooth or rough. This metadata matters wherever appearance carries information, such as retail catalogues, industrial imagery and recognition datasets in which similar-looking objects are made of different things.
MaterialTextureFinishTransparency
ExampleTable – wooden, polished. Tool – metallic. Window – transparent glass. Wall – rough concrete.
08 · Environmental context
Environmental Context Annotation
Context annotation records the conditions around the main subjects: site type, terrain, ground condition, weather, vegetation, nearby structures and infrastructure. The same machine in a warehouse, on a construction site and in a field is a different data point. Context tags let you filter, balance and evaluate a dataset by those conditions instead of discovering the gaps after training.
Site typeTerrainWeatherVegetationInfrastructure
ExampleConstruction site, dry soil, overcast weather, nearby vegetation, industrial surroundings.