SelvaBox / README.md
CanopyRSAdmin's picture
Update README.md
c1e5208 verified
|
raw
history blame
2.23 kB
metadata
dataset_info:
  features:
    - name: height
      dtype: int64
    - name: width
      dtype: int64
    - name: fold
      dtype: string
    - name: raster_name
      dtype: string
    - name: location
      dtype: string
    - name: image
      dtype: image
    - name: tile_name
      dtype: string
    - name: annotations
      struct:
        - name: bbox
          sequence:
            sequence: float64
        - name: segmentation
          dtype: 'null'
        - name: area
          sequence: float64
        - name: iscrowd
          sequence: int64
        - name: is_rle_format
          dtype: 'null'
        - name: category
          sequence: string
    - name: tile_metadata
      struct:
        - name: crs
          dtype: string
        - name: transform
          sequence: float64
        - name: bounds
          sequence: float64
        - name: width
          dtype: int64
        - name: height
          dtype: int64
        - name: count
          dtype: int64
        - name: dtypes
          sequence: string
        - name: nodata
          dtype: 'null'
  splits:
    - name: train
      num_bytes: 16884458976
      num_examples: 585
    - name: validation
      num_bytes: 2786536280
      num_examples: 387
    - name: test
      num_bytes: 11729504363.402
      num_examples: 1477
  download_size: 31336873232
  dataset_size: 31400499619.402
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: cc-by-4.0

SelvaBox: A high-resolution dataset for tropical tree

This is the version of the SelvaBox dataset that has been pre-processed and presented in our SelvaBox paper.

The dataset is made of 14 rasters resampled at 4.5 cm GSD, from three different countries: Brazil, Ecuador and Panama. It comprises over 83 000 human bounding box annotations for tropical tree crowns in dense canopies.

Training tiles are 3555x3555 pixels, while validation and test tiles are 1777x1777 pixels, equivalent to 80x80 meters spatial extent. There is 50% overlap between train and validation tiles, and 75% between test tiles (to ensure that the largest trees of 50+ meters in diameter will fit entirely in at least one tile).

Link to our paper will come soon.