Glossary

Glossary

Terms as vidlabel uses them. Each links to the page that defines it fully.

term	meaning
annotation	One row in a video's annotation table: a marker, a box or a mask, with a label and provenance. Concepts
annotation table	The Parquet file videos/<video_id>/annotations.parquet. One row per annotation, all kinds together. Spec
author	Column naming the person who made or reviewed a row.
box	A rectangle on one frame, stored as pixel corners x0, y0, x1, y1 with x1 and y1 exclusive.
build	Compile the per-video tables into one sorted, content-hashed release. Releases
chunk	In a ReadPlan: the requested frames of one video that fall in one keyframe interval, read in one pass. Training
codec (mask)	How a mask is stored in bytes. The only one is bitpack_zstd. Masks
compiled release	See release.
confidence	Optional number from 0 to 1 a model gave a row.
content hash	sha256:<hex> of a video file; how a dataset recognises "the same file".
dataset	A folder in the vidlabel layout. Concepts
deep validation	validate --deep: also decodes every mask and checks it against its box.
extra	A JSON-object column (and metadata field) for your own data under your own key.
frame	A 0-based index in display order. end_frame is inclusive.
frame cache	Optional, derived JPEG or PNG copies of frames so random access is a file read. Not part of the dataset. Training
frame index	Optional frame_index.json: a timestamp for every frame and the keyframe numbers. Spec
keyframe	A frame that can be decoded without earlier frames. Random access costs a decode from the keyframe before the frame wanted.
kind	marker, box or mask.
label	A name from vocab.json.
marker	A span of time with no geometry (start_frame to end_frame).
mask	A boolean pixel region on one frame, stored cropped to its bounding box, bit-packed and zstd-compressed.
origin	Who produced a row: human, model, imported, other, or your own x-... value.
ReadPlan	An arrangement of reads for a list of frames you chose, so each keyframe interval is decoded once. Training
release	compiled/<release_id>/: the whole dataset merged into one sorted table plus a vocabulary snapshot. The id is a hash of the file. Releases
rotation	Clockwise angle (0, 90, 180, 270) applied to the stored frame for display. Coordinates are in the displayed orientation.
spec	The format definition: docs/SPEC.md and the vidlabel-spec package.
status	unreviewed, approved or rejected. Rejected rows stay in the file but are left out of releases.
tool, tool_version	Which program (and version) produced a row.
track	A track_id shared by the rows of one object across frames, scoped to one video and view. Concepts
view	A named crop and/or set of time ranges of one video, optionally resized. Rows carry a view_id tag; their coordinates stay full-frame. Views
vocab	vocab.json: the one shared list of labels.
writer	Whoever modifies a video's annotation file. Rule today: one writer per video at a time.

---

Source: docs/glossary.md