abstreet/docs/TODO_quality.md

TODO - Map quality

Geometry

• try fixed pt again, for determinism purposes mostly

  • very different approaches

    • v1: keep f64 internally, but constantly drop small bits
      • we just want deterministic serialization really, so...
    • v2: try some rational number thing internally

  • retry integers for everything (probably all at once)

    • dont round too early... line intersection, dist to pt, kinematics
    • adjust epsilon values to match actual smallest difference, or start doing direct comparisons now

  • clamp distances first, not points? This one is working well!

    • note contains_pt needs to use 2 or 3 * epsilon, because of the error that may accumulate...
    • audit all EPSILON_DIST usages
    • make natively orderable

  • change internal pt2d representation to int. JUST get that working first.

    • then get rid of approx_eq, approx_dedupe
    • make sure dist_to at the tiniest case rounds up, too. maybe that needs to happen now.
    • make Pt2D natively orderable, hashable
    • can Pt2D::new() and x() and y() return something besides f64?

  • then work on proper SI types, with negative/positive cases handled carefully

  • also bounds?

  • cant get rid of the ccw intersection check... different answer in some cases that looks bad

• underlying problems

  • bad polyline shifting remains
    • from the remaining cases, looks like we need to totally remove some tight points and retry
    • make polygons use the corrections too?

• generalized_trim_back

  • breaks down when we have jagged lane endings due to polyline shift angle correction
  • sometimes a lane polyline hits the perpendicular of a trimmed road! where was this happening?

• handle small roads again somehow?

  • VERY overeager... ate half of the map

    • can we capture snapshots of incremental changes?
      • save initial map at every step, be able to load raw + initial with a focus point
      • generic viewer should be easy... something that stores polygon and ID, wraps the quadtree, etc
    • try merging the shortest roads first

  • deal with loop roads

  • manually draw a picture of the weird intersection to see what would look reasonable. i think we need original road bands from deleted stuff to make decent polygons.

  • model U-turns

• degenerate-2's should only have one crosswalk

  • then make them thinner

• ped paths through sidewalk corners are totally broken

  • calculate the better paths first, then make the corner geometry from that?

• car turns often clip sidewalk corners now

• figure out what to do about yellow center lines

  • intersections on one-ways look weird
  • yellow and white lines intersect cars and turn icons and such
  • who should own drawing them?

More data

• lanes: https://data-seattlecitygis.opendata.arcgis.com/datasets/49d417979fec452981a068ca078e7070_3
  • not filled out for most streets
• traffic circles: https://data-seattlecitygis.opendata.arcgis.com/datasets/717b10434d4945658355eba78b66971a_6
• https://data-seattlecitygis.opendata.arcgis.com/datasets/sidewalks
  • disagrees with OSM road centers sometimes
• https://data-seattlecitygis.opendata.arcgis.com/datasets/curb-ramps
• high quality thick roads: https://seattlecitygis.maps.arcgis.com/apps/webappviewer/index.html?id=86cb6824307c4d63b8e180ebcff58ce2
• OSM has footways
  • but theyre not marked everywhere
  • and theyre hard to associate with roads (sometimes need to infer a planter strip)
• draw ALL water and greenery areas
• draw benches, bike racks
• render trees
• look for current stop sign priorities
  • https://gis-kingcounty.opendata.arcgis.com/datasets/traffic-signs--sign-point/

Low-priority geometry issues

• if building front path intersects another building, then scrap that building.

  • or wait, just require bldgs to be even closer to sidewalk first.
  • need to do polygon vs polygon check!
  • will need to speed it up with quadtree containing entire buildings. make sure these are easy to use.

• can we make OSM buildings with holes?

  • experiment with https://docs.rs/clipping/0.1.1/clipping/gh/struct.CPolygon.html and https://github.com/21re/rust-geo-booleanop

More problems to fix

• Disconnected map
  • Some lane-changing model needed to determine this
• Impossible turns (from a far bus lane to a crazy left)

Release

• publish the map data
• stabilize the code except for sim