DSTL Competition. Examples of automatic neural net segmentation
Competition: https://www.kaggle.com/c/dstl-satelli...
Details of solution: https://www.kaggle.com/c/dstl-satelli...
This is the visualization of how trained convolutional neural net segment maps in fully automated way. Images generated with solution, which got 2nd place in Dstl Satellite Imagery Feature Detection contest on Kaggle.
There were 10 classes to segment images: 1- Buildings, 2 - Misc structures, 3 - Roads, 4 - Tracks, 5 - Trees, 6 - Crops, 7 - Fast water, 8 - Slow water, 9 - Trucks, 10 - Small cars
First part of video (00:00 - 00:39) shows how provided picture in full size segmented for different classes.
Second part of video (00:39 - 01:14) shows how small region from full size picture (3360x3360 pixels) segmented for different classes.
Third part of video (01:14 - 02:17) compare real segmentation on the left, with predicted segmentation on the right. The pictures are from validation set, e.g. it didn't take part in CNN training process.
Last part of video (2:17 - 2:42) the same as third part but for small subregion of full size picture.