ilteralp / remote_sensing_with_gcn Goto Github PK

Images will be segmented into superpixels with using alpha and omega trees where each superpixel is treated as a graph node. To use multi-scale cues, multiple graphs will be generated using different alpha/omega values.

This idea is from the work of Wan et al. They used edge hops to create multi-scales. At the scale s, every superpixel x_i is connected to its s-hop neighbors.

Classification results from the model are not reproducible and learning through epochs is not stable.

Experimented with setting seed, causes every epoch to have the same accuracy.

Instead of representing whole image as a graph, multiple graphs will be generated from the image and classification will be done using these graphs.

Number of classes in data object returns number of nodes.

Train and test accuracy does not change through epochs.

The alpha and omega values mentioned in issue #5 will be learned by the model.