Solution to the datascience challenge contest Put it on the map!

link to the contest: https://www.xstarter.io/challenge-details/put-it-on-the-map

Goal: retrieve map lattitude and longitude coordinates

Database: 40k images with known coordinates for training and 10k images for testing

Example maps (different map-styles, resolutions and annotations):

Download the training and test data from the contest website

Workflow of the proposed solution:

• encoding images using transfert learning from pretrained ResNet model on a geographic zones classification task (Figure 1 shown above)
• binary hash images of training images in 512 feature space
• retrieving the closest image to the one to identify in that feature space using Locality Sensitive Hashing for fast approximate nearest neighbor

Set up the environment

Run:

$ pip install -r requirements.txt

Training

Run the notebook Find_similar_images_Training_submissionFinal.ipynb to access the solution

Inference

Run the notebook Find_similar_images_Inference_submissionFinal.ipynb to apply the trained models to the test database

Interesting feature: Ability of the model to find closest images in the database robust to resolutions, annotations, image motions and map types

3 random examples of closest images:

Once your model is trained run several the last cell of Find_similar_images_Training_submissionFinal.ipynb notebook to see more random examples.