Outputs¶

Body Eye Sync writes results as Parquet files so they can be loaded directly by Python analysis tools such as pandas, Polars, or PyArrow.

Main Output¶

Each input receives one main output:

outputs/<input-id>.parquet

The core tracking columns are:

Column	Meaning
`frame`	Zero-based video frame index
`track_id`	Stable track identifier
`x1`, `y1`, `x2`, `y2`	Tracked object box in video pixels
`conf`	Object detection confidence

When face detection is enabled, face columns are merged onto matching (frame, track_id) rows:

face_score
face_x1, face_y1, face_x2, face_y2
left_eye_x, left_eye_y
right_eye_x, right_eye_y
nose_x, nose_y
mouth_left_x, mouth_left_y
mouth_right_x, mouth_right_y

When body-pose detection is enabled, pose columns are merged onto matching rows:

pose_score
pose_x1, pose_y1, pose_x2, pose_y2
Per-keypoint pose_<name>_x, pose_<name>_y, and pose_<name>_score columns for the COCO keypoints.

Embedding Outputs¶

If embeddings are collected, companion files are written beside the main output:

outputs/<input-id>.body_embeddings.parquet
outputs/<input-id>.face_embeddings.parquet

Embedding files contain:

Column	Meaning
`track_id`	Track the embedding belongs to
`frame`	Frame the embedding came from
`score`	Detection score used for top-k selection
`embedding`	Fixed-size float16 vector

Only the best embeddings_per_track vectors are kept for each tracklet.

Reading Results¶

import pandas as pd

tracks = pd.read_parquet("outputs/camera-1.parquet")
body_embeddings = pd.read_parquet("outputs/camera-1.body_embeddings.parquet")