Pexels-Pairs-Masklets-330K (review sample)

Per-video masklet annotations stored as Parquet shards.

This repository is a small sample released for anonymous peer review. It contains 5 shards drawn from 5 different set_* directories of the full collection, which holds roughly 330K shards across 301 sets. Contents are unmodified; only the number of shards is reduced.

Layout

set_0000/<video_id>.mp4.parquet
set_0075/<video_id>.mp4.parquet
set_0150/<video_id>.mp4.parquet
set_0225/<video_id>.mp4.parquet
set_0300/<video_id>.mp4.parquet

One Parquet file per source video. Each row is a sampled frame with its object masklets.

Schema

column	description
frame_idx	index of the sampled frame within the video
frame_image	encoded frame image
frame_height, frame_width	frame dimensions in pixels
native_fps, target_fps	source frame rate and the rate used for sampling
objects	per-object masklet records for the frame
size	source video file size
nb_frames	total frames in the source video
r_frame_rate	source rational frame rate
duration	source video duration in seconds

Usage

import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

p = hf_hub_download(
    "anonymousML123/Pexels-Pairs-Masklets-330K",
    "set_0000/VAD-WEHFg64-1080.mp4.parquet",
    repo_type="dataset",
)
table = pq.read_table(p)
print(table.schema)