Downsampled Disentanglement Datasets - Falcor3D and Isaac3D
Anonymous · Figshare · 2020
New disentanglement datasets This data repository contains the Falcor3D and Isaac3D datasets for disentanglement learning, where the image resolution is 128x128. Falcor3D The Falcor3D dataset consists of 233,280 images based on the 3D scene of a living room. The meta code corresponds to all possible combinations of 7 factors of variation: lighting_intensity (5) lighting_x-dir (6) lighting_y-dir (6) lighting_z-dir (6) camera_x-pos (6) camera_y-pos (6) camera_z-pos (6) Note that the number m behind each factor represents that the factor has m possible values, uniformly sampled in the normalized range of variations [0, 1]. Each image has as filename padded_index.png where index = lighting_intensity * 46656 + lighting_x-dir * 7776 + lighting_y-dir * 1296 + lighting_z-dir * 216 + camera_x-pos * 36 + camera_y-pos * 6 + camera_z-pos padded_index = index padded with zeros such that it has 6 digits. Isaac3D The Isaac3D dataset consists of 737,280 images, based on the 3D scene of a kitchen. The meta code corresponds to all possible combinations of 9 factors of variation: object_shape (3) object_scale (4) camera_height (4) robot_x-movement (8) robot_y-movement (5) lighting_intensity (4) lighting_y-dir (6) object_color (4) wall_color (4) Similarly, the number m behind each factor represents that the factor has m possible values, uniformly sampled in the normalized range of variations [0, 1]. Each image has as filename padded_index.png where index = object_shape * 245760 + object_scale * 30720 + camera_height * 6144 + robot_x-movement * 1536 + robot_y-movement * 384 + lighting_intensity * 96 + lighting_y-dir * 16 + object_color * 4 + wall color padded_index = index padded with zeros such that it has 6 digits.