Raw images or Mappillary images

[Rub21]

@karitotp, once we get the AOI from https://github.com/developmentseed/housing-passports-v2/issues/3, could you make an evaluation of which image we should use? the raw images or Mapillary, Also, could you determine which attributes have been cleaned by Mapillary from the original images? It looks like we need to ensure that we have enough attributes to make the inferences and attach them to the building footprints. cc @piligab @yunica

[karitotp]

To find out the attributes of the raw images, we generated a CSV file with the attributes of the sample images from Road_S1_SPL, in total there are more than 100 attributes per image.

On the other hand, to find out which attributes of raw images are kept in Mapillary and which are not, we wanted to upload some images from the sample folder (that contains Road_S1_SJH and Road_S1_SPL folders) that World Bank shared with us, but we noticed that the images had already been uploaded to Mapillary (those Mapillary images appear with the captured date of Aug 15, 2023.), the images are from Dominica, as we can see in the image below.

Having properties of both sources we have been doing some comparisons and so far what we have noticed is the following:

1. The attributes from Mapillary that remain the same values as the raw image properties are: altitude, captured_at, height, and width.
2. All other attributes from the raw images are removed or transformed once they are uploaded to Mapillary to the following properties:
   • atomic_scale
   • camera_type
   • compass_angle
   • computed_altitude
   • computed_compass_angle
   • computed_rotation
   • exif_orientation
   • merge_cc
   • mesh
   • sequence
   • sfm_cluster

@yunica, it would be good if you could review the properties of the raw images, to know if they can help us to attach images with the building footprints.

Here we share:

• The CSV file, that was uploaded it on this spreadsheet
• Mapillary points that were obtained from a part of the Road_S1_SPL

cc. @piligab

[Rub21]

From the previous project phase, the required metadata are following attributes:

heading[deg],image_fname,frame,latitude[deg],longitude[deg],cam,neighborhood,subfolder
238.81604587884,ladybug_14062044_20190628_141526_Cube_000000_Cam1_879_090-0881.jpg,ladybug_14062044_20190628_141526_Cube_000000,-0.937739981443431,100.377668449777,"1",padang,PADANG_01
238.81604587884,ladybug_14062044_20190628_141526_Cube_000000_Cam3_879_090-0881.jpg,ladybug_14062044_20190628_141526_Cube_000000,-0.937739981443431,100.377668449777,"3",padang,PADANG_01
238.918032366373,ladybug_14062044_20190628_141526_Cube_000001_Cam1_880_091-3497.jpg,ladybug_14062044_20190628_141526_Cube_000001,-0.93774492395104,100.377660678983,"1",padang,PADANG_01

I think mapillary has the attribute heading[deg] as compass_angle,

            "properties": {
                "captured_at": 1692120982000,
                "compass_angle": 304.014,
                "id": 300097322676782,
                "is_pano": true,
                "organization_id": 276431331814934,
                "sequence_id": "jLa60AtxBSz2d1Ey87bcFO"
            }

Also talking with Junior we can get higher resolution images apart of what the team found here: https://github.com/developmentseed/housing-passports-v2/issues/6#issuecomment-1753570569

Conclusions:

@srmsoumya we are going to use mapillary images for labeling and inference.

cc. @geohacker @developmentseed/data-team

[karitotp]

Also talking with Junior we can get higher resolution images apart of what the team found here: https://github.com/developmentseed/housing-passports-v2/issues/6#issuecomment-1753570569

We obtained the new Mapillary images with the improvements in the resolution, and they look much better than those we got previously. So, with these improvements in the image resolution, we can work with Mapillary images.

Here an example of the new Mapillary clipped images and raw clipped images in 1024x1024

New Mapillary image 1024x1024

Raw image 1024x1024

cc. @Rub21 @yunica

[srmsoumya]

Thanks @Rub21 @karitotp for looking into this.