We’ve rolled out a new quality_score field on Mapillary images and sequences. We would love feedback from folks that consume data from Mapillary APIs.
It’s a float in the range [0, 1] representing the predicted visual quality of an image or a sequence — higher is better. The intent is to make it easy to detect blurry, dark, or otherwise low quality captures.
How to query it — images (Graph API)
Add quality_score to the fields= parameter on any image-returning endpoint:
GET https://graph.mapillary.com/{image_id}?fields=id,quality_score&access_token=MLY|…
{
“id”: “2140279232804782”,
“quality_score”: 0.704
}
Image endpoints that expose it:
- GET /{image_id} — single image
- GET /images?bbox=… — bulk by bounding box
- GET /images?sequence_ids=… — images in a sequence
- GET /images?organization_id=… — images from an organization
- GET /{map_feature_id}/images — images depicting a map feature
Combine with other fields as usual, e.g. fields=id,quality_score,captured_at,geometry.
How to query it — sequences (vector tiles)
There is no Graph API entity endpoint for a sequence’s quality_score — it lives only on the vector tile sequence layer, as the average of per-image scores in that sequence (unscored images excluded). Fetch the relevant coverage tile and read the field off each sequence feature:
- https://tiles.mapillary.com/maps/vtp/mly1_public/2/{z}/{x}/{y}
- sequence layer (zoom 6-14) — sequence-level quality_score (average)
- image layer (zoom 14) — per-image quality_score
- https://tiles.mapillary.com/maps/vtp/mly1_computed_public/2/{z}/{x}/{y}
- same two layers, on the SfM-computed coordinates
Quick Python sample (using mapbox_vector_tile):
import requests, mapbox_vector_tile
raw = requests.get(
“https://tiles.mapillary.com/maps/vtp/mly1_public/2/14/8801/5374”,
params={“access_token”: “MLY|…”},
).content
tile = mapbox_vector_tile.decode(raw)
for f in tile[“sequence”][“features”]:
print(f[“properties”][“id”], f[“properties”].get(“quality_score”))
Not included: the overview layer (zoom 0-5) on either coverage tileset, and the map-feature tilesets (mly_map_feature_point, mly_map_feature_traffic_sign) — those are dedup objects, not images.
Notes
- Type: nullable float in [0, 1]
- Images that haven’t been scored yet may come back as either 0 or null — handle both
- On the sequence tile layer the field is absent when none of the images in the sequence have a score yet
Things we’d like to hear about:
- Does the quality score values make sense ?
- Any endpoints where you expected to see it but don’t?
- Anything else — bugs, gaps etc
Drop your thoughts 
to use. 6 
clicks to report one image. Plus, there is no way to report a whole sequence. Actual effects of these reports are also rather mixed at best.
Every camera (or picture) is only as good as the photographer behind it. The same goes for pencils and writers. 
Though, I admit that would the platform drop less than 0.5 quality score images over night, it may be sobering or cathartic to some people. 
The image quality score and the reporting function are not meant for shaming, judging, or grading people but tools to keep the database clean, which is desperately needed for every openly publicly (non-gated) fed database.
