search

mapbox / robosat

Semantic segmentation on aerial and satellite imagery. Extracts features such as: buildings, parking lots, roads, water, clouds
MIT License
2.01k stars 382 forks source link

poor prediction in satellite images #219

Open jaigsingla opened 3 years ago

jaigsingla commented 3 years ago

Hi All,

I am able to get 0.83 of IOU using the Tanzania dataset using your code and able to generate vector files after predictions.
Now, I have a 50cm VHR image over an urban area and corresponding building shape files in geojson format. I applied the same techniques as in the Tanzania dataset and trained the model with an IOU of .67 But, generated segmentation probabilities do not look good, what can be the reason ? I have used batch_size of 16, image_size of 512, lr= 0.00005 and loss function of Lovasz. I will be happy to hear from you.

regards, Jai

daniel-j-h commented 3 years ago

Can be anything from

By now this project is outdated, unmaintained, and no longer state of the art, too.

On June 7, 2021 9:07:17 AM UTC, jaigsingla @.***> wrote:

Hi All,

I am able to get 0.83 of IOU using the Tanzania dataset using your code and able to generate vector files after predictions.
Now, I have a 50cm VHR image over an urban area and corresponding building shape files in geojson format. I applied the same techniques as in the Tanzania dataset and trained the model with an IOU of .67 But, generated segmentation probabilities do not look good, what can be the reason ? I have used batch_size of 16, image_size of 512, lr= 0.00005 and loss function of Lovasz. I will be happy to hear from you.

regards, Jai

jaigsingla commented 3 years ago

Hi Daniel,

You are right it may be anything. Can you guide me through? It will be preety useful. In case you have commercial software, pls let me know.

On Fri, 18 Jun 2021, 18:21 Daniel J. H., @.***> wrote:

Can be anything from

  • not enough data
  • poor optimizer settings (lr too low?)
  • not trained for long enough
  • trained for too long
  • ..

By now this project is outdated, unmaintained, and no longer state of the art, too.

On June 7, 2021 9:07:17 AM UTC, jaigsingla @.***> wrote:

Hi All,

I am able to get 0.83 of IOU using the Tanzania dataset using your code and able to generate vector files after predictions. Now, I have a 50cm VHR image over an urban area and corresponding building shape files in geojson format. I applied the same techniques as in the Tanzania dataset and trained the model with an IOU of .67 But, generated segmentation probabilities do not look good, what can be the reason ? I have used batch_size of 16, image_size of 512, lr= 0.00005 and loss function of Lovasz. I will be happy to hear from you.

regards, Jai

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/mapbox/robosat/issues/219#issuecomment-864015048, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGFXXGFB73D353FUHISBII3TTM6M7ANCNFSM46HGVQ4Q .

manapshymyr-OB commented 3 years ago

Hi All,

I am able to get 0.83 of IOU using the Tanzania dataset using your code and able to generate vector files after predictions. Now, I have a 50cm VHR image over an urban area and corresponding building shape files in geojson format. I applied the same techniques as in the Tanzania dataset and trained the model with an IOU of .67 But, generated segmentation probabilities do not look good, what can be the reason ? I have used batch_size of 16, image_size of 512, lr= 0.00005 and loss function of Lovasz. I will be happy to hear from you.

regards, Jai

Hello! on which OS have you performed this project?

manapshymyr-OB commented 3 years ago

Can be anything from - not enough data - poor optimizer settings (lr too low?) - not trained for long enough - trained for too long - .. By now this project is outdated, unmaintained, and no longer state of the art, too. … On June 7, 2021 9:07:17 AM UTC, jaigsingla @.***> wrote: Hi All, I am able to get 0.83 of IOU using the Tanzania dataset using your code and able to generate vector files after predictions. Now, I have a 50cm VHR image over an urban area and corresponding building shape files in geojson format. I applied the same techniques as in the Tanzania dataset and trained the model with an IOU of .67 But, generated segmentation probabilities do not look good, what can be the reason ? I have used batch_size of 16, image_size of 512, lr= 0.00005 and loss function of Lovasz. I will be happy to hear from you. regards, Jai

Hello! Is it possible to run on OS Windows without Docker?