Building Inspector - Shape + Address consensus

Transcript

1. mauricio giraldo arteaga @mgiraldo NYPL Labs
3. bon jour
4. my name is mauricio
6. research and circulating library system spanning the Bronx, Staten Island and Manhattan boroughs in NYC
8. NYPL Labs
10. i’m going to talk about maps
11. The Great Map Data Extraction
12. an adventure in three acts and a prologue and an epilogue
13. prologue
14. The Lionel Pincus and Princess Firyal Map Division
21. 500,000+ maps 20,000+ books & atlases
27. year
28. street names year
29. use type street names year
30. use type street names name year
31. material use type street names name year
32. material use type street names name class year
33. material use type street names address name class year
34. material use type street names address floors name class year
35. material use type street names address floors name class year skylights
36. material use type street names address floors name class year skylights backyards
37. material use type street names address floors name class geo location year skylights backyards
38. footprint material use type street names address floors name class geo location year skylights backyards
39. footprint material use type street names address floors name class geo location year skylights backyards
40. we got these for several decades since the 1800s and by 1950 every town in the US with a population of 2,000 had been mapped
41. data trapped in a legacy format
42. we want all the data!
43. f**k yeah historical data!
44. citysdk.waag.org/buildings
45. citysdk.waag.org/buildings
46. NYU Stern / Imaginaria3D
47. NYU Stern / Imaginaria3D
48. maps.google.com
49. maps.google.com
51. data
52. it all starts with a photograph
54. but it is “just a photo” but it is only a few clicks away
56. maps.nypl.org/warper
59. geo-rectification or: “make it match Open Street Map”
62. *this is a simulation. actual process is intensive. consult your mathematician before trying
65. vectorization or: “draw the building shapes”
67. results from maps.nypl.org/warper
68. hand-crafted, artisanal, locally-sourced data
69. 500,000+ maps 20,000+ books & atlases
70. 500,000+ maps 20,000+ books & atlases* *imagine how many pages an atlas has
71. in the order of dozens of millions building footprints if counting NYC only
73. ~120k footprints produced in three years
    by staff and volunteers
75. this will take us a few millenia* *actual number taken out from a hat
76. there has to be a better way
77. act i: will there be polygons?
78. requests to geo companies went unanswered
80. can we automate this?
82. ¿¡quoi!? @mgiraldo
87. what is a building?
89. completely enclosed by black lines
90. completely enclosed by black lines dashed lines are not walls
91. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2)
92. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2)
93. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2) not paper-colored
94. process
95. github.com/NYPL/map-vectorizer
100. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2) not paper-colored
101. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2) not paper-colored
102. provide the best (possible) input image
107. differences in resampling cubic nearest neighbor
108. differences in resampling cubic nearest neighbor
109. make the image a binary bitmap or: “black and white”
112. polygonize or: “convert contiguous pixels to a single line shape”
114. ! gdal_polygonize.py test.tif -f "ESRI Shapefile" test.shp test
115. ! gdal_polygonize.py test.tif -f "ESRI Shapefile" test.shp test
116. ! gdal_polygonize.py test.tif -f "ESRI Shapefile" test.shp test
117. ! gdal_polygonize.py test.tif -f "ESRI Shapefile" test.shp test
119. no no no no no
120. no no no no no yes yes
121. simplify* *for those polygons that we care about
122. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2) not paper-colored ✔ ✔
125. alpha shape *code basically stolen wholesale from rpubs.com/geospacedman/alphasimple
126. ﹡ ﹡ ﹡ ﹡ ﹡﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
127. ﹡ ﹡ ﹡ ﹡ ﹡﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
128. ﹡ ﹡ ﹡ ﹡ ﹡﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
129. we need a set of points
131. pts = spsample(polygon, n=1000, type="hexagonal")
132. pts = spsample(polygon, n=1000, type="regular")
133. pts = spsample(polygon, n=1000, type="random")
134. now we alpha shaping
135. x.as = ashape(pts@coords, alpha=2.0)
136. x.as = ashape(pts@coords, alpha=2.0)
137. x.as = ashape(pts@coords, alpha=2.0)
138. there are other point reduction algorithms like Ramer-Douglas-Peucker or Whyatt Curve Simplification
139. separate the buildings from the chaff
140. completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2) not paper-colored ✔ ✔ ✔ ✔
143. [218, 211, 209]
144. [218, 211, 209] paper [199, 179, 173], [179, 155, 157], [206, 193, 189], [199, 195, 163], [207, 204, 179], [195, 189, 154], [209, 203, 181], [255, 225, 40], [194, 198, 192], [161, 175, 190], [137, 174, 163], [166, 176, 172], [149, 156, 141] [205, 200, 186] not paper
148. this is good enough for our use case
152. ✔ ✔ ✔ ✔ ✔ completely enclosed by black lines dashed lines are not walls > 20m2 (~180ft2) < 3,000m2 (~27,000ft2) not paper-colored
153. computer-vision for attribute recognition *bonus quest
157. 66,056 footprints produced in one day for an 1859 atlas of Manhattan
158. caveats:
     ! adjacency not enforced
     false positives/negatives
     buildings may also overlap
159. act ii: the vectorizer needs to prove itself
164. multiple inspections for each item and let consensus surface on its own
165. footprint validation or: “tell us what the computer got right or wrong“
166. are people willing to spend time checking building footprints? insurance atlases are not exactly the coolest type of maps
168. buildinginspector.nypl.org
169. github.com/NYPL/building-inspector
174. about a month later…
179. 420k+ flags* 70k+ unique polygons ! consensus: ~84% YES, 7% FIX, 9% NO *a “flag” is a YES/NO/FIX by one person for a given polygon
180. seems people are willing after all… we — our contributors
181. seems people are willing after all… we — our contributors
182. act iii: the return of the inspector
183. footprint material use type street names address floors name class geo location year skylights backyards
184. divide and conquer
185. footprint material use type street names address floors name class geo location year skylights backyards
186. three new tasks for now… we really want it all!
188. footprint material use type street names address floors name class geo location year skylights backyards
189. check
190. check YES
191. check YES address color
192. check YES FIX address color
193. check YES FIX address color fix
194. check YES FIX address color fix
195. check YES FIX address color fix *footprints marked as “NO” go to building heaven
196. check YES FIX address color fix *footprints marked as “NO” go to building heaven
197. fix
198. fix
199. address
200. address
201. classify color
202. classify color
203. 865k+ flags
204. check YES FIX address color fix
205. check YES FIX address color fix for 80k+ unique polygons 77k+ 5k+ 42k+ 26k+
206. epilogue
207. address and shape consensus or: how to determine what the right building footprint and address looks like?
210. all points are useful inclusiveness above all
219. DBSCAN for the win citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.71.1980
220. bit.ly/nypl-consensus
221. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
222. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
223. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
224. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
225. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
226. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
227. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
228. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
229. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ + +
230. 246 246 246 414 246 414 414 246 414 414 414 ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ + +
231. 246 246 246 414 246 414 414 246 414 414 414 ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ + +
232. 246 414 + +
235. DBSCAN for shapes also!
242. all points are still useful
244. ﹡
245. ﹡ ﹡
246. ﹡ ﹡ ﹡
247. ﹡ ﹡ ﹡ ﹡
248. ﹡ ﹡ ﹡
249. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
250. ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡ ﹡
251. + + + + + + +
252. + + + + + + +
253. + + + + + + +
254. + + + + + + +
255. + + + + + + +
256. + + + + + + +
257. + + + + + + +
258. + + + + + + +
266. resulting data available via an API
267. resulting data available via an API in 100% recyclable GeoJSON
269. photographing
270. photographing ↓
271. photographing ↓ geo-rectification
272. photographing ↓ geo-rectification ↓
273. photographing ↓ geo-rectification ↓ vectorization
274. photographing ↓ geo-rectification ↓ vectorization ↓
275. photographing ↓ geo-rectification ↓ vectorization ↓ inspection
276. photographing ↓ geo-rectification ↓ vectorization ↓ inspection ↓
277. photographing ↓ geo-rectification ↓ vectorization ↓ inspection ↓ check / fix / color / address
278. photographing ↓ geo-rectification ↓ vectorization ↓ inspection ↓ check / fix / color / address ↓
279. photographing ↓ geo-rectification ↓ vectorization ↓ inspection ↓ check / fix / color / address ↓ consensus
280. photographing ↓ geo-rectification ↓ vectorization ↓ inspection ↓ check / fix / color / address ↓ consensus ↓
281. photographing ↓ geo-rectification ↓ vectorization ↓ inspection ↓ check / fix / color / address ↓ consensus ↓ data release
282. not the end
286. ¡merci beaucoup! mauricio giraldo arteaga @mgiraldo NYPL Labs slides at: bit.ly/nypl-ehess images from: NYPL digital collections - Wikimedia Commons
 Christopher Cannon - Flickr user wallyg - Giphy