We recently noticed that a traffic sign we had downloaded a few months ago was no longer showing up in the Mapillary map nor in the API.
Below is the example, where you can see an 80km/h and a 60km/h traffic sign identified in the photos (taken from 2022). The 80km/h traffic sign was amongst the signs downloaded, but when we went back and check it had disappeared from the map and also from the API, but still showing up as identified in the image. Actually, the 60 one still shows up but from older imagery. If you filter the map for 2022 as start date, it won’t show up either in the map.
I don’t know how widespread this is, but it’s a very big deal for us who are trying to build traffic sign inventories for a few cities. We ran a comparison of the traffic signs dataset we had downloaded for Rio de Janeiro back in May and found that the new one had 234 traffic signs less than the one downloaded 2 months ago. Here is another example where you can see a bunch of traffic signs in the pictures, but not showing up in the map any longer (they were in the older dataset we downloaded).
@boris or @tao not sure if you had already identified this, but if not, for you to be aware.
@boris@abalys and @tao following up on this. When querying the map features endpoint using their MAP_FEATURE_ID, how should we interpret a response for an object which no longer returns any geometry fields, like the one below:
This is the CURL curl --location --max-time 100 'https://graph.mapillary.com/546477557876758?access_token=XX&fields=id%2Cgeometry%2Cobject_value%2Cobject_type'
Does this mean the traffic sign was taken out from the map for some reason or why are the coordinates no longer being returned (fyi, this was a traffic sign we downloaded from the API a few weeks ago, but stop showing up in the map and in the API)?
Objects have missing geometries if they are not fully processed. Object placement on the map is first computed from imagery of each upload and then recomputed by using other imagery in the area, which in this case had not happened when you were posting your message.
Geometries also change when new imagery is introduced in the area to produce more accurate results. My hypothesis is that new imagery is added to the area and objects get marked for reprocessing, which takes time.
Longer term(I believe mid next year), we will improve processing to be much faster and users will get better feedback on when everything is processed.
Are you sure? Because from what I have observed lately, objects (traffic signs and feature points) have not been consolidated any longer for years. If anything, objects on the map are now only produced based on the oldest sequence where an object has been detected first. Detections from later (more recent) sequences are not considered for the object map layer any longer (they are still stored in the database) and not used as additional triangulation data for consolidation with existing (old) objects. You can verify this by clicking on an object on the map layer and looking at the list of images that have the object detected. No new images are added to this list from any recently uploaded sequence where the same object has been detected.
Sometimes, we see the same detected (phyiscal) objects spilled multiple times from different sequences in the same area on the map layer. They are not consolidated either. Maybe this is only the case if the sequences in question all have different projection parameters imagery (or types of imagery; planar vs. spherical)? This also did not use to be an issue in the past when objects were consolidated. So, I do not know but the current object triangulation and consolidation situation is a bit of a mess.