README.md

metadata

dataset_info:
  - config_name: default
    features:
      - name: image
        dtype: image
      - name: depth
        dtype: image
      - name: label
        dtype: image
    splits:
      - name: train
        num_bytes: 1829417279.14
        num_examples: 5285
      - name: test
        num_bytes: 1747976639.6
        num_examples: 5050
    download_size: 2452649738
    dataset_size: 3577393918.74
  - config_name: uint8
    features:
      - name: image
        dtype:
          image:
            mode: RGB
      - name: depth
        dtype:
          image:
            mode: L
      - name: label
        dtype:
          image:
            mode: L
    splits:
      - name: train
        num_bytes: 673574397.52
        num_examples: 5285
      - name: test
        num_bytes: 598216510.95
        num_examples: 5050
    download_size: 916719066
    dataset_size: 1271790908.47
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
  - config_name: uint8
    data_files:
      - split: train
        path: uint8/train-*
      - split: test
        path: uint8/test-*

SUN RGB-D

Easier version for semantic segmentation.

default config contains RGB and uint16 version of depth images.

uint8 config contains RGB and uint8 version of depth images, I convert uint16 by divide the pixel values by 255 and save it in new depth image.

Comes from:

SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite (CVPR 2015) (PDF) (Website)