UWGZQ/VidSTG-SAM3.1-Masks

VidSTG SAM 3.1 object masks (train)

SAM 3.1 mask tracks for the relation objects of VidSTG train videos (VidOR videos), one row per object per frame with a mask. Masks come from vidstg-masks (commit 72c28f8, default settings, prompted from the VidOR boxes). No video frames are included.

rows (masks)	13,880,398
videos	5,119
objects	19,195 (79 categories)

Columns

column	type	meaning
video_id	string	VidOR / VidSTG video id
object_id	int32	VidOR object id (tid) within the video
frame_id	int32	native-fps frame index, the VidOR annotation's index space
category	string	VidOR category of the object
height, width	int32	mask size in pixels
mask_rle	string	COCO compressed RLE of the mask

Rows are sorted by video_id, object_id, frame_id. A frame where an object has no mask (absent, or SAM gave no usable mask) has no row.

import pandas as pd
from pycocotools import mask as mask_utils

df = pd.read_parquet("vidstg_train_sam31_masks.parquet")
r = df.iloc[0]
m = mask_utils.decode({"size": [r.height, r.width], "counts": r.mask_rle.encode()})  # H x W uint8

Coverage

All 5,436 VidSTG train videos were processed; 5,119 have masks here. Not included:

reason	videos
VP6-encoded source video (removed)	207
ran out of GPU memory (very long clips, > ~17.5k frames x objects)	102
more than 16 objects (not run by the pipeline)	8

Sources: 4,791 videos from two runs of ours and 328 from salabajr/vidstg-masks-train500 (revision 64a3d8a), same pipeline commit.