finalvent/michiyomi-tokyo-streetscape
michiyomi — Tokyo streetscape verbalization open data English Overview michiyomi pairs coordinates with structured Japanese descriptions of physical streetscapes visible in public Mapillary imagery. A vision-language model (VLM) verbalized only what is visible in each image: no map, address, place name, facility name, statistics, or other external knowledge was injected. Release 2026-09-13-r1 contains 1,914,490 scenes covering all of Tokyo: the 23… See the full description on the dataset page: https://huggingface.co/datasets/finalvent/michiyomi-tokyo-streetscape.
michiyomi — Tokyo streetscape verbalization open data
English
Overview
michiyomi pairs coordinates with structured Japanese descriptions of physical streetscapes visible in public Mapillary imagery. A vision-language model (VLM) verbalized only what is visible in each image: no map, address, place name, facility name, statistics, or other external knowledge was injected. Release 2026-09-13-r1 contains 1,914,490 scenes covering all of Tokyo: the 23 wards (now with multiple capture years), the Tama area (26 cities, 3 towns, 1 village) and the islands (2 towns, 7 villages), the latter two mainly at their latest capture year.
Images themselves are not redistributed. Each scene’s id is a string Mapillary image ID that links the record back to its source image.
Coverage warning: All of Tokyo is now included, but coverage is not uniform street-by-street. A scene is a representative image per grid × capture direction × year, so density follows where contributed Mapillary imagery exists. A missing record means “not covered by this dataset,” not “the feature does not exist.” Use the ward breakdown below or the API’s coverage endpoint before drawing conclusions about a location.Release history
The dataset metadata remains language: ja because the record content—summary, structured analysis, change category, subject, evidence, and parts of verification metadata—is Japanese. Interface documentation is bilingual.
How the data was produced
- Source material is public street imagery from Mapillary, captured by citizens and organizations. The images are not included in this dataset.
- Each image was verbalized from visual evidence only. The output describes what was visible from that camera angle in that capture year; it is not a place or facility identification.
- Longitudinal changes in
changesbegan as hypotheses from images of the same location in different years. A separate, context-isolated session from the same model family attempted to falsify each claim. Only claims that survived withstatus="supported"are included. Shared systematic error can remain, so the supported rate is not an accuracy guarantee. schoolscontains official elementary-school locations from Japan’s National Land Numerical Information P29. It is not derived from imagery and is intended as a coordinate-discovery aid.
Before using the data
- Always state the capture year. A description from a 2018 image reflects the image’s 2018 view, not current conditions.
- Descriptions include monocular-image estimates. Width and similar numeric values carry confidence information and are not survey measurements.
- The data preserves “unknown” and “unknown outside the frame.” Do not fill in facts that the record does not contain.
- Thirty-nine anomalous capture-time records remain in the dataset with
quarantined=1(including epoch-anomalous timestamps such as 1970); they are excluded from nearby search and coverage statistics in the online API but remain retrievable by ID. segment_priority_scoreinchangesis a road-segment inspection-priority heuristic, not the danger or severity of a change; it is null for many rows.- Keep Mapillary image IDs as strings. Converting them to numbers can lose precision.
- If you translate Japanese narratives, retain the Japanese source text and record the translation method, model/version, date, and source-content hash.
Quickstart
from datasets import load_dataset
scenes = load_dataset(
"finalvent/michiyomi-tokyo-streetscape",
"scenes",
split="train",
)
print(scenes[0]["id"]) # string Mapillary image ID
print(scenes[0]["capture_year"]) # year represented by the image
print(scenes[0]["summary"]) # Japanese summaryLoad the smaller tables by replacing "scenes" with "changes", "schools", or "generations". For bulk spatial analysis, read the ward-partitioned Parquet files directly. For online coordinate queries, use the public API.
Tables and schema
scenes — 1,914,490 rows, partitioned by municipality (ward / city / town / village)
One row represents one scene: a representative image for a grid × capture direction × year. Three data layers are joined in each row.
Municipality-partition counts: ota 128,016 / setagaya 117,106 / minato 105,188 / nerima 96,029 / koto 90,332 / shibuya 76,633 / hachioji 76,254 / shinjuku 71,609 / chiyoda 68,424 / shinagawa 60,631 / chuo 60,255 / edogawa 57,505 / suginami 55,840 / taito 55,044 / bunkyo 53,462 / machida 47,336 / toshima 47,310 / itabashi 44,952 / adachi 44,948 / sumida 40,179 / meguro 34,907 / kita 31,567 / katsushika 29,770 / fuchu 29,457 / chofu 25,601 / nakano 20,450 / tachikawa 20,119 / inagi 19,641 / mitaka 19,446 / arakawa 18,731 / tama 17,930 / higashimurayama 17,766 / kunitachi 16,459 / kodaira 14,307 / musashino 14,076 / hino 13,730 / ome 12,915 / okutama 12,724 / kokubunji 11,627 / oshima 11,371 / hinohara 9,795 / akishima 9,413 / nishitokyo 9,405 / kiyose 8,264 / niijima 8,220 / koganei 8,165 / mizuho 8,100 / akiruno 7,311 / musashimurayama 7,299 / higashikurume 6,481 / fussa 5,768 / higashiyamato 5,024 / miyake 4,449 / outside23 4,222 / komae 4,083 / kozushima 3,881 / ogasawara 3,660 / hamura 3,213 / hachijo 2,524 / aogashima 2,283 / hinode 1,736 / toshima-island 1,544 / mikurajima 3.
- Of the 1,914,490 rows, 1,899,979 use SfM-corrected Mapillary coordinates; raw GPS is used only when corrected coordinates are unavailable.
position_sourcemakes the choice explicit. - Machine-layer computation ran for every scene, but individual fields can remain unavailable when inputs are missing (for example, travel bearing 99.6% and absolute objects 86.2%). Color statistics are available for all rows except 39 flagged
color_status="error:TypeError". The API’s/v1/metais the canonical source for field coverage. - Capture years span 1986–2026 for non-quarantined rows, and 99.8% of scenes are from 2014 onward; 3,660 non-quarantined scenes predate 2014 (the 23-ward past-year coverage added in v0.3 reaches back further than v0.2). Records before 2014, and especially before 2010, are more likely to carry clock anomalies; only epoch-level anomalies are quarantined.
- Updates are published as versioned releases. API responses include
snapshotso the queried release can be identified later.
changes — 2,522 rows
Contains only longitudinal-change claims that survived adjudication. Fields include Japanese category (equipment, building, road marking, pavement, roadside use, vegetation, or other), year_a → year_b, Japanese subject, full Japanese evidence, and verification_json. Coordinates are road-group centers, not exact changed-object locations.
schools — 1,323 rows
Elementary-school locations in Tokyo from National Land Numerical Information P29 (unchanged from v0.2).
generations — generation provenance
Both generations in this release passed adoption gates and are released.
Generation and prompt versions are traceable through per-row provenance columns. Each scene belongs to one served generation.
API and MCP
The public API and MCP server serve the same release as this dataset. Use the dataset for bulk analysis and the API for coordinate-based online queries.
- REST API: https://michiyomi.dev/v1 — unauthenticated, read-only nearby search, coverage, scene detail, changes, and catalog endpoints.
- English OpenAPI 3.1: https://michiyomi.dev/openapi.en.yaml
- Japanese OpenAPI 3.1: https://michiyomi.dev/openapi.yaml
- English developer guide: https://michiyomi.dev/en/docs/
- MCP server:
https://michiyomi.dev/mcp— tools:find_school,coverage,describe_location,get_scene, andchanges_near. - Browser UI: https://michiyomi.dev/en/
The API’s /v1/meta release ID matches this dataset: 2026-09-13-r1. The API serves 1,914,451 of the 1,914,490 rows in nearby search and coverage (quarantined rows are excluded there but retrievable by ID).
License and attribution
CC BY-SA 4.0. Share-alike applies to derived works because the data is derived from Mapillary. Display this attribution when using the data:
Derived from © Mapillary contributors (CC BY-SA 4.0) / Derived from MLIT National Land Numerical Information (A29 and P29) / Tokyo Metropolitan Government Bureau of Construction emergency transport roads
A UI displaying source imagery must show a visible Mapillary logo and link to https://www.mapillary.com, in accordance with Mapillary’s terms.
Known limitations
- All of Tokyo is included, but scene density is uneven and follows where contributed Mapillary imagery exists; the Tama area and the islands are mostly single-year.
- Cloud/clear-sky boundary classifications can vary. A systematic gen1 error can misclassify elevated expressways as ordinary roads; gen2 was calibrated for this issue.
- A 27B-class model does not maintain perfect self-consistency. Deterministic checks detect and repair some contradictions, while remaining issues are traceable through provenance columns.
- Photographer identity and camera make/model are not included in the public dataset.
Citation
michiyomi — Tokyo streetscape verbalization open data (release 2026-09-13-r1)
https://michiyomi.dev/ / Derived from Mapillary contributors (CC BY-SA 4.0)日本語
概要
東京都(23区・多摩地域・島しょ)の街路画像(Mapillary)を Vision-Language Model が画像に写っているものだけから言語化した、「座標 × 構造化テキスト」のデータセットです。release 2026-09-13-r1 の1,914,490シーン(東京都全域。23区は複数年、多摩地域・島しょは最新年中心)を収録しています。レコードの本文は日本語のため、YAMLメタデータの language は ja のままです。
収録範囲の注意: 東京都全域を収録していますが(23区は複数年、多摩・島しょは最新年中心)、街路単位の被覆は一様ではありません。シーンは「グリッド × 撮影方位 × 年」の代表画像なので、密度は市民が投稿したMapillary画像の分布に従います。レコードがないことは、現地に対象が存在しないことを意味しません。区別内訳またはAPIの coverage で被覆を確認してください。リリース履歴
データの成り立ち
- 素材は Mapillary の街路画像(市民・事業者が撮影して公開したもの)。画像そのものは配布しません。各行の
idが Mapillary image ID で、原典画像は Mapillary で参照できます。 - 各画像をVLMが視覚情報のみから言語化しています。地図・住所・施設名などの外部知識は一切注入していません。したがって記述は「その画像に写っていたもの」であり、地名や施設の同定結果ではありません。
- 経年変化(
changes)は、同一地点の異なる年の画像を比較して生成した仮説を、文脈分離した別セッションによる反証優先の再検証にかけ、反証しきれなかったもの(status="supported")だけを収録しています。検出側と検証側は同じモデル系列であり、共有された系統誤差は残りえます。 - 小学校位置(
schools)は国土数値情報(P29)由来の公的データで、画像由来ではありません。座標探索の起点として使えます。
読む前に知っておくべきこと(正直性の規約)
- 撮影年に注意。 各シーンには
capture_yearが付いています。2018年の画像は2018年の状態であって現在ではありません。 - 記述は単眼画像からの推定を含みます。幅員などの数値にはconfidenceが付いており、実測値ではありません。
- 「不明」「画角外不明」はそのまま「不明」として収録しています。このデータにないことを補完・創作しないでください。
- 撮影時刻が異常なシーンは
quarantined=1でフラグしています(28件。1970年などのエポック異常を含みます。行は残してあります)。 segment_priority_score(changes)は道路区間の点検優先度ヒューリスティックであり、変化そのものの危険度ではありません(多くの行でnull)。- Mapillary image IDは文字列です。数値に変換しないでください(桁落ちします)。
- 日本語記述を翻訳するときは、原文を保持し、翻訳手法・モデル/版・日時・原文ハッシュを記録してください。
クイックスタート
from datasets import load_dataset
scenes = load_dataset(
"finalvent/michiyomi-tokyo-streetscape",
"scenes",
split="train",
)
print(scenes[0]["id"]) # 文字列のMapillary image ID
print(scenes[0]["capture_year"]) # この画像が表す撮影年
print(scenes[0]["summary"]) # 日本語の要約テーブル構成
scenes(既定・1,914,490行、市区町村別Parquet)
1行 = 1シーン(グリッド × 撮影方位 × 年の代表画像1枚)。3つの層を結合済みです。
区別ファイル: setagaya 89,408 / ota 84,491 / minato 70,314 / nerima 66,008 / koto 61,493 / shibuya 54,263 / shinjuku 54,161 / chiyoda 46,999 / suginami 41,853 / shinagawa 41,054 / taito 40,659 / edogawa 40,169 / chuo 39,409 / bunkyo 38,214 / toshima 35,571 / itabashi 33,459 / adachi 32,464 / sumida 29,710 / meguro 27,639 / kita 24,861 / katsushika 20,611 / nakano 17,198 / arakawa 15,020 / outside23 4,222
- 処理総数は1,914,490件です。撮影時刻異常の39件は
quarantined=1のままデータセットに保持し、APIの近傍検索・被覆統計から除外しています(ID直接参照は可能)。 - 座標はSfM補正座標(computed)を原則採用し(1,899,979件)、欠損時のみ生GPSを使用します。採用元は
position_sourceに明示します。 - 機械層は全シーンで計算済みですが、入力欠損により項目ごとに未充足があります(travelbearing 99.7%・absobjects 88.0%など)。色統計は32件のみ
color_status="error:TypeError"で評価不能を明示しています。充足率の正本はAPIの/v1/metaです。 - 撮影年は検疫外の行で1986〜2026年、99.97%は2014年以降です(2013年以前は332シーン)。2014年以前、特に2010年以前の行は時計異常の可能性が相対的に高く、エポック級の異常のみ検疫しています。
- 更新はversionedなreleaseとして不定期に行います。API応答の
snapshotで問い合わせ対象の版を特定できます。
changes(2,522行)
裁定を生き残った経年変化のみ。category(設備/建物/区画線・標示/舗装/沿道用途/植栽/その他)、year_a→year_b、subject、evidence、verification_json(裁定の全文)を収録します。座標は道路グループ中心であり変化地点そのものではありません。
schools(1,323行)
東京都の小学校位置(国土数値情報P29。v0.2から変更なし)。
generations(世代来歴)
言語化世代とモデルの対応表。本リリースは2世代とも検証済み(released)です。
世代・プロンプト版数は各行の来歴列で判別できます。1シーン1配信世代です。
API・MCP
このデータセットと同一のリリースを配信する公開APIとMCPサーバがあります。バルク分析はこのデータセット、座標からのオンライン検索はAPIが向いています。
- REST API: https://michiyomi.dev/v1 — 認証なし・read-only。
- 日本語OpenAPI: https://michiyomi.dev/openapi.yaml
- 英語OpenAPI: https://michiyomi.dev/openapi.en.yaml
- 開発者ガイド: 日本語 / English
- MCPサーバ:
https://michiyomi.dev/mcp—find_school/coverage/describe_location/get_scene/changes_near。 - ブラウザUI: https://michiyomi.dev/
APIの /v1/meta が返すリリースIDとこのデータセットのリリースIDは一致します(2026-09-13-r1)。APIの近傍検索・被覆は1,914,490件のうち検疫外の1,914,451件を対象にします(検疫行もID直接参照は可能)。
ライセンスと出典表示
CC BY-SA 4.0。派生物にも継承されます(Mapillary由来のため)。利用時は次の出典を表示してください。
© Mapillary contributors (CC BY-SA 4.0) を加工 / 国土交通省 国土数値情報(A29・P29)を加工 / 東京都建設局 緊急輸送道路
画像を表示するUIにはMapillaryロゴの視認可能な表示と https://www.mapillary.com へのリンクが必要です(Mapillary利用規約)。
既知の制約
- 東京都全域を収録していますが、シーン密度は投稿されたMapillary画像の分布に従い一様ではありません。多摩・島しょはほぼ単年です。
- 晴/曇の境界判定にブレがあります。高架専用道の
facility_typeはgen1世代に「一般道路」への誤分類の系統誤差が確認されています(gen2では較正済み)。 - 27Bクラスのモデルは自己整合を完全には維持できません。内部矛盾は決定論チェッカーで検出・リペアし、残余は来歴列から追跡できます。
- 撮影者情報・カメラ機材情報は本データセットに含めていません(内部来歴としてのみ保持)。
引用
michiyomi — 東京都 街路言語化オープンデータ (release 2026-09-13-r1)
https://michiyomi.dev/ / Mapillary contributors (CC BY-SA 4.0) を加工