BARISTA

BARISTA is a densely annotated egocentric video dataset of coffee preparation, designed for unified benchmarking of vision-language models across spatial, temporal, relational, and procedural understanding tasks.

The dataset contains 185 egocentric videos (~4.4 hours, 30 FPS, 1280×720 to 1920×1080) covering three coffee preparation methods: capsule machines, portafilter machines, and fully automatic machines. Videos were recorded in controlled indoor setups using iPhones, Apple Vision Pro, RayBan Meta 3, and RayBan Wayfarer smart glasses.

Dataset structure

Each video is stored in its own directory:

<video_id>/
    coco_annotation.json   # COCO-style annotations (masks, bboxes, attributes, relations, activities)
    video.mp4              # raw video

coco_annotation.json follows the COCO format extended with additional top-level keys:

Key	Description
annotations	Per-frame instance annotations. Fields: id, image_id (0-based frame index), object_id (UUID), bbox ([x, y, w, h]), segmentation (COCO RLE with counts and size), area
attributes	Segment-level key-value attributes per object. Fields: id, object_id, attribute_type (e.g. "color", "state"), value, image_ranges (list of {image_id_start, image_id_end})
relations	Directed typed relations between object pairs. Fields: id, source_object_id, target_object_id, relation_type (e.g. "position", "human_actions"), value, image_ranges
categories	Object categories. Fields: id (UUID), name.
activities	Fine-grained verb+noun activity segments. Fields: id, display_name, activity_class_id (UUID), image_range ({image_id_start, image_id_end})
process_steps	High-level process step segments. Same fields as activities
object_id_to_category_id	Map from object UUID to category UUID (needed to resolve annotation object_id → category)
video_metadata	List with one entry. Fields: document_id, video_index, width, height, frame_count, fps, length_in_ms, recording_device_type, recording_device_version
split	Dataset split: "train" or "test"

Loading the data and running evaluations

See the project repository for the dataset loader and the VLM benchmarking pipeline.