Dronescapes (2023)

This is the original Dataset, released in 2023 in our work presented at ICCV: link. It consists of 10 UAV videos, split in 7 videos used for training the models and 3 for test set as follows:

TODO: remake this table

In total there are 556 human annotated frames for semantic segmentation, out of which 116 are used for the test set. The other 440 human annotated frames were propagated using SegProp. This method propagates the semanatic segmentation labels between two human annotated frames through optical flow across all the missing frames. In total, it resulted in all 23,568 frames having a semantic label.

The dronescapes classes

• TODO: present them and color map

• TODO: present distribution (for weights)

• TODO: depth/camera normals and sfm