---
title: Hard Intersection Multimodal Sample
emoji: 🚦
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
startup_duration_timeout: 30m
pinned: false
short_description: Multimodal AV dataset of a hard Tokyo intersection
---

# Hard Intersection Multimodal Sample

This Space runs the open source FiftyOne App in a Hugging Face Docker Space.

The Space loads
[`Voxel51/hard-intersection-multimodal-sample`](https://huggingface.co/datasets/Voxel51/hard-intersection-multimodal-sample)
into local ephemeral storage before FiftyOne starts. The dataset is
downloaded again whenever Hugging Face provisions a fresh container.

This is a **grouped** dataset: each of the 4 groups is one driving episode
through the Takanawadai intersection in Tokyo, Japan, with 6 synchronized
camera video slices and 1 3D point cloud slice per group. Frame-level labels
include COCO semantic detections, Lanelet2 HD map overlays, and the vehicle
trajectory projected onto the front camera.

Each browser receives a temporary clone of the dataset records while all
clones share the same read-only media. Filters, modal selection, playback,
tags, fields, and saved views therefore remain isolated between visitors.
Inactive clones expire after 30 minutes, cleanup runs every five minutes,
and the Space allows at most 20 active browser sessions.

## Use your own datasets

Duplicate this Space and edit [`datasets.json`](./datasets.json) to replace or
extend the dataset list. Each entry must point to a FiftyOne-formatted Hub
dataset and an absolute local download path, then rebuild.