Reading the annotations
An annotation describes an image at three levels: individual visible elements, semantic groups and their parent links, and selected spatial relationships. The complete schema defines required fields and allowed values; DATA_FORMAT.md describes the packaged fields.
Elements and groups
| Node kind | Types | Purpose |
|---|---|---|
| Element | text, image, shape |
Individual visible regions |
| Group | chart, axis, legend, legend_item, data_item, series, panel |
Regions that organize related children |
Every explicit node has a label, parent, bbox, desc, color_palette, and node_kind. Its type is recorded in element_type or group_type, depending on its kind. Each label is unique within one chart; it is not a global dataset identifier. Pair labels with sample_id when joining annotations from multiple images.
parent encodes the hierarchy and refers to an existing label or ROOT. The root is implicit: it is counted in dataset statistics but is not another object in the nodes list. For N explicit nodes, there are N hierarchy links and N + 1 nodes including the root. Do not count a group as a visual element or add a second explicit root when calculating these totals.
Semantic roles
Element roles distinguish visual function from rendering type. A text element can be a heading, tick label, value label, or source credit; an image or shape can also serve a data-related function.
| Role family | Roles |
|---|---|
data_expression |
data_mark, value_label, direct_readout |
reading_reference |
axis_line, tick_mark, tick_label, axis_title, gridline, legend_key, legend_label, legend_title, entity_label, reference_line, annotation_connector |
context |
heading, explanation, source_credit, subject_depiction |
presentation |
boundary, separator, emphasis, ornament |
Recognized visible wording belongs in a text node's text field. desc provides an appearance and reconstruction description. Keep both fields when using annotations: a description is not a substitute for recognized text.
Boxes and spatial relationships
Dataset boxes use [y0, x0, y1, x1], with integer coordinates normalized to 0–1000. Multiply horizontal coordinates by width / 1000 and vertical coordinates by height / 1000 to obtain pixels. These are axis-aligned logical boxes, not segmentation masks.
The separate spatial_relations list uses node labels as endpoints and stores selected sibling-pair relationships across the four spatial types. The original selected records, their order, types, endpoints and distance fields are unchanged; is_near is added to the existing distance records. Hierarchy links are stored separately through parent. An unrecorded pair is not a negative label.
is_near is a Boolean on each stored bbox_distance record. Positive-area overlap and containment are false; exact edge or corner contact is true. Otherwise the shortest box gap must be at most both 0.06 × min(width, height) and 1.0 × min(sqrt(area_a), sqrt(area_b)). Compute gaps and areas in pixels using the actual image width and height. The Boolean uses unrounded geometry; distance_ratio remains gap divided by the image diagonal.
Example and limitations
The included training example contains the exact image, graph object, spatial list, and row identity from the dataset. It illustrates the file format and is not an independent annotation-quality measurement.
Annotations are produced by the dataset annotation pipeline. The 1,000-chart test split, comprising 500 real and 500 synthetic charts, has been manually verified as described in the paper's annotation-quality assessment. The human_gold field is true for test records and false for training records.
Image2SceneGraph's native predictions use a different layout field and box order. Consult the model's output-format guide before comparing a prediction to this dataset.