Split the 360 images and get left and right images

[Rub21]

@srmsoumya , As we discuss, we are going to split in to chunks the 360 images, we can use the tools the we built for CPAL. https://github.com/developmentseed/spherical2images#readme,

@piligab Could you run this process,? and upload the images into different folder(right and left) in the same bucket s3://hp-images-v2

cc. @karitotp @yunica

[piligab]

We have clipped the raw images of the sample images - Road_S1_SPL. They are available on s3 --> s3://hp-images-v2/sample/Road_S1_SPL_clip_images/

Afterward, we reviewed the clipped images and we found the following:

1. Image size:

We have explored the image size (1024x1024 and 512x512) in the clip images, so we noticed that image size 1024x1024 has better visualization than 512x512 size images, as we can see in the images below.

512x512 pixels

1024x1024 pixels

2. Image quality - Mapillary vs Raw:

Comparing the images of Mapillary and Raw in 1024x1024, it is clearly noticeable that the quality of Mapillary is lower than Raw images, this low resolution of Mapillary images will make it difficult to recognize and labeling of characteristics like building material and building condition. Here are some examples:

Example 1: Mapillary image 1024x1024

Raw image 1024x1024

Example 2: Mapillary image 1024x1024

Raw image 1024x1024

So, according to our findings, it will be better if we use the Raw images with 1024x1024 pixels.

cc. @srmsoumya, @Rub21, @yunica, @karitotp

[Rub21]

Thank you, @piligab, for providing the information. @srmsoumya, What do you think the resolution.? Do you think it's something we should consider when training the module? i think the team may use the raw images..

[piligab]

Verified the resolution of the images improved on this ticket, we are using the Mapillary images. So we already obtained the Mapillary points from the whole Dominica with the following filter: Organization ID and Timestamp from Jan 01, 2023, getting a total of 18,693 Mapillary points.

On the other hand, according to Nuala's answer about those images outside the AOI provided, we will work with all available images, that is, we incorporate the additional points into the AOI initially provided.

Then, we did some processing on the points, to remove those repeated points, this processing is basically the same as we did in the CPAL project, simplifying the sequences and Mapillary points by distance. For instance, the results that we obtained are:

• Total Mapillary points by sequence filter (buffer=0.000010): 14,179 points.
• Simplified Mapillary points per 2-meter distance: 13,486 points

So, after running this simplification process, in the end, we have a total of 13,486 points, and considering that each point is an image that is going to be clipped to get its left and right side, we would have 26, 972 images for labeling.

cc. @srmsoumya @Rub21 @yunica @karitotp

[piligab]

The data are available on s3:

• Mapillary points (initial file) --> s3://hp-images-v2/mapillary_files/mapillary_points_panoramic.geojson
• Mapillary points (final file after processing) --> s3://hp-images-v2/mapillary_files/map_points__pano_filtered_simplify_2.geojson
• Mapillary points (clipped file) --> s3://hp-images-v2/mapillary_files/map_points__pano_clip.geojson
• Mapillary images --> s3://hp-images-v2/mapillary_images/