by Muhammad Shalahuddin Yahya Sunarko

To see the original README.md of this repo, click here.

Citation of Sat2Graph:

He, S., Bastani, F., Jagwani, S., Alizadeh, M., Balakrishnan, H., Chawla, S., … Sadeghi, A. (2020). Sat2Graph: Road Graph Extraction through Graph-Tensor Encoding. arXiv [Cs.CV]. Retrieved from http://arxiv.org/abs/2007.09547

1. Architecture: Sat2Graph
2. Dataset: Omani Cities, satellite imageries from Mapbox, road graph networks from OpenStreetMap. We have 161 regions (train:val:test = 0.7:0.1:0.2). But, during training, only 0.35 out of 0.7 were used because of memory issue (from the author’s original training script).
3. Total training duration: about 48 hours (224200 steps) on Quadro RTX 4000
4. Best model: step 174000, test_loss 0.579
5. Tile size: 352x352

No	Satellite imagery	GT of Semantic Seg	Predicted Graph	Predicted Semantic Seg
1
2
3
4
5
6
7
8

1. The model performs well detecting roads, even those that are missed to be annotated by humans. See segmentation GT in (4) and (6), we found some roads are not annotated in the GT, but our model is able to correctly detect that. In (7) and (8), we see that the model is barely can detect unannotated roads. Maybe because Sat2Graph is not just using simple segmentation for doing graph prediction, but it also accounts the road network topology and connectivity so is robust even though there are several missing and misaligned roads in the ground truth.
2. The model still has some false negatives.

1. Solve memory limitation issue, so we can train the model using full data/higher number of data
2. Gather more data from other cities, in hope that the model can better generalize and less affected by missing roads
3. Filter out training images with incorrect annotations (high effort)
4. Add road networks dataset from targeted cities