Awesome

HydroViet_VOR

Visual Object Retrieval in satellite images with Triplet Network generated from Template (Processing)

Data Structure

Train data and validation data contain 5 classes: dams, buildings, house, river, bridge. Collect by hands from Google Earth (5000 images for now)
Test data contains 100 images (for now)

HydroViet_VOR
|
└───train_data
│   └───house
│       └───house1.jpg
│       └───house2.jpg
│       └───...
│   └───river
│       └───...
|   └───...
|
└───val_data
│   └───house
│       └───house1.jpg
│       └───house2.jpg
│       └───...
│   └───river
│       └───...
|   └───...
|
└───test
|   │   1.jpg
|   │   2.jpg
|   |   ...

Method

Using Pairwise Ranking Loss model
Find the distance between 2 feature vector in embedding space.
Loss function is MarginRankingLoss (use Euclidean distance to calculate loss)
Image below is borrowed from https://gombru.github.io/2019/04/03/ranking_loss/

model

Train

Config hyperparameters in configs/config.py

python trainv1.py

Test

Visualize embeddings (test_data) using tSNE (clustering)

python test.py

Visualize

Using cosine similarity to plot most similar images base on the query image

python visualize.py

Result

The best accuracy when validating is 56%, which is not good enough

Furthur Improvement

Using GANs to train CNN models like EfficientNet, Resnet on satellite data

Sources

Make sure to give them a star!

Template: https://github.com/kaylode/custom-template/
Supervised Contrast Learning: https://github.com/HobbitLong/SupContrast/