Mapillary’s computer vision technology automatically detects objects depicted in images, such as buildings, cars, pedestrians, bicycle racks, and much more. This is based on a method called semantic segmentation—an algorithm is trained to detect and assign a category label to every pixel in the image.
Since each image has geocoordinates, it is possible to also visualize the object detections on the map. This will let you filter for images that contain specific kinds of objects. The map visualizations at this stage reflect positions of the images where certain objects are visible. This is not the same as map features, which is one step further—triangulating the location of an object that has been detected in several images, so we can determine the location of that object in the real world and place it on the map as a map feature.
You can visualize object detections on the Mapillary web app. (See the article about exploring object detections.)
Please note that to retrieve and use these automatically extracted object detections, you must query map data using the Mapillary API.
Below is the list of the 65 classes of objects that are provided in addition to traffic signs and point features.
| Curb Fence Guard rail Barrier (other) Lane separator Wall Bike lane Crosswalk (basic) Curb cut Parking Pedestrian area Rail track Road Road shoulder Service lane Sidewalk Traffic island Bridge Building Garage Tunnel Person |
Lane marking (dashed) Lane marking (solid) Crosswalk (zebra) Lane marking (other) Lane marking (stop line) Lane marking (text) Sky Snow Lawn Vegetation Water Banner Bench Bike rack Billboard Catch basin Fire hydrant Junction box Mailbox Manhole Parking meter Pothole |
Streetlight |
Comments
8 comments
What is the principle of triangulating the location of detected objects in several images and retrieving the real world location of objects?
Hi! Wikipedia has a great article on triangulation: https://en.wikipedia.org/wiki/Triangulation_(computer_vision)
This is the principle we use as well when triangulating object detections, in order to generate map features that represent the objects' real-world locations.
Trees are listed but do not show up. Vegetation in general is not listed. Need for analysis of greencity environment.
Hi! "Trees" are actually "vegetation" and we've apparently forgotten to update the documentation here. Fixed now and thank you for pointing this out.
Hi! Unless I am missing something, I can not find "Wire group" category label in "mapillary-vistas-dataset_public_v2.0.zip" dataset. Am I missing something? Thanks in advance!
Hello, it seems that we can't see anymore the semantic segmentation in the web app, I think you should update this article
is the object detection automatic and instant after uploading images to mapillary (obv after the ingesting), or does it take more time to complete or something to do?
En el caso de las señales graficadas en el mapa, existen algunas que están repetidas como se pueden filtrar los recorridos? Luego al descargar la información de estas señales, puedo hacerlo en coordenadas cartesianas (Este, Norte)?
Please sign in to leave a comment.