SpatialSemanticDataset

SpatialSemanticDataset is a dataset of first-person walking videos annotated with camera poses and English semantic descriptions. It organizes annotations by scene and supports research on visual navigation, camera trajectory and language alignment, and descriptions of observations and intended actions.

The dataset covers 12,079 scenes from 3,188 videos across 9 source channels. This repository contains the scene index and annotations. The source videos are hosted separately and can be located by their YouTube video IDs.

The data processing pipeline is maintained in a separate repository. SpatialSemanticPipeline

Annotation methodology

The accompanying pipeline uses TransNetV2 to detect scene boundaries in source videos. It runs VGGT-Omega on overlapping video chunks to estimate camera positions and orientations. The pipeline checks trajectory quality, and the released pose CSVs include a status field. Rows marked bad can contain missing pose values.

A vision-language model generates timestamped English annotations from video clips. When trajectory-derived navigation hints are available, the pipeline includes them in the prompt. Otherwise, it uses a prompt without hints. Each annotation contains an Observation describing visible content, a Latent Thought expressing the model's inference about intent or reasoning, and an Action describing movement.

See SpatialSemanticPipeline for the implementation, configuration, and prompt templates.

Repository layout

Path	Contents
metadata.csv	Scene index, with one row per scene
poses/<channel>/<video_id>/<scene_id>.csv	Timestamped camera positions, orientations, and pose status
semantics/<channel>/<video_id>/<scene_id>.txt	Timestamped English observations, inferences, and action descriptions

Use (channel, video_id, scene_id) to join the index with the two annotation files. Scene IDs such as scene_0000 are local to each video.

Data fields

Scene index

Field	Meaning
channel	Source channel
video_id	YouTube video ID
scene_id	Scene identifier within the source video
start_second, end_second	Scene start and end times in the source video, in seconds

Camera poses

Field	Meaning
source_t	Timestamp in the source video, in seconds
source_frame	Frame index in the source video
status	Pose status. Sample files contain valid rows with pose values and bad rows whose pose fields can be NaN.
x, y, z	Camera position components
qx, qy, qz, qw	Camera orientation quaternion, with the scalar component qw last

The dataset files leave the position units, axis conventions, and coordinate relationships between scenes unspecified. Confirm these conventions before calculating physical distances or combining trajectories across scenes.

Semantic annotations

Annotations are grouped by source-video timestamp in [HH:MM:SS] format. Each block contains an Observation describing what is visible, a Latent Thought expressing an inferred situation or intent in the first person, and an Action describing an action or plan.

Pose samples and semantic annotations can occur at different times. Join them by scene, then align them on the source-video timeline.

Load a scene

Install huggingface_hub to download the public dataset. This example downloads the index and the two annotation files for its first scene.

import csv
from pathlib import Path

from huggingface_hub import hf_hub_download

REPO_ID = "MonsterDie/SpatialSemanticDataset"


def download(filename):
    return Path(hf_hub_download(REPO_ID, filename, repo_type="dataset", token=False))


with download("metadata.csv").open(encoding="utf-8-sig", newline="") as file:
    scene = next(csv.DictReader(file))

key = "/".join(scene[field] for field in ("channel", "video_id", "scene_id"))

with download(f"poses/{key}.csv").open(encoding="utf-8-sig", newline="") as file:
    valid_poses = [row for row in csv.DictReader(file) if row["status"] == "valid"]

annotations = download(f"semantics/{key}.txt").read_text(encoding="utf-8-sig")

print(scene)
print(f"Valid poses: {len(valid_poses)}")
print(annotations.split("\n\n", 1)[0])

The dataset ships without predefined splits. When creating training and evaluation splits, group scenes by (channel, video_id) so that scenes from the same source video stay in the same split.

License

The dataset card declares the MIT license. Source videos are subject to their own terms.