UIBK Avalanche Dataset

UIBK Avalanche Dataset

Dataset of 4090 labelled avalanche photographs. Each photograph in the dataset was assigned to one of four categories: glide, loose, slab, or none. These labels refer to the different avalanche release mechanisms: glide, loose-snow, and slab, with none indicating the absence of avalanches. In addition,  avalanche experts outlined each visible avalanche with a polygonal bounding box and assigned it a single label of glide, loose, or slab. 

The photographs were taken in the field in the winters of 2000/2001 - 2021/2022 and each image has the date and approximate location at which it was taken in the image name. Most photographs in the dataset were taken from the ground: additional images from low-flying helicopters were included only if the annotators deemed that they could feasibly have been taken by webcams.

Accompanying code for training ResNet, VGG, and YOLOv3 models to detect avalanches is available at github.com/j-f-ox/avalanche-detection

Folders

• annotations      # Polygonal and rectangular bounding boxes for each avalanche image
  • glide
  • loose
  • slab
• images             # The .jpg images sorted by overall image label
  • glide
  • loose
  • none
  • slab
• test.txt              # Test images
• train.txt             # Training images

Label Distribution 

• The images folder contains 716 glide, 416 loose, and 1887 slab avalanche photographs and a further 1071 "none" images without visible avalanche regions
• The dataset contains a total of 2489 annotated glide avalanches, 1827 loose-snow avalanches, and 2912 slab avalanches