Custom Dataset

Module containing the custom torch.utils.data.Dataset classes used in this package.

class master_thesis.dataset.MasterThesisDataset(bgs_dataset_meta, masks_dataset_meta, bgs_simulator, masks_simulator, split, **kwargs)

Bases: torch.utils.data.dataset.Dataset

Implementation of the custom dataset class used in the package.

Both the background and the mask will be either extracted from a real sequence or simulated using single real frames after applying affine transformations. The CLI arguments --p-simulator-bgs and -p-simulator-masks define the probability that a certain item data item has been obtained using the simulation mode.

If self.kwargs['frames_n'] is set to -1, the entire sequence associated with the desired item is returned.

bgs_dataset

Instance of the dataset returning background images.

masks_dataset

Instance of the dataset returning binary masks.

split

Split of the dataset.

kwargs

Dictionary containing the CLI arguments used in the execution.

FILL_COLOR = tensor([0.4850, 0.4560, 0.4060])
__getitem__(item)

Returns an item from the dataset, where an item is composed by a background image, its masked version, the binary mask and some auxiliary information.

Parameters

item – Index of the data item to return.

Returns

Tuple of the form ((x, m), y, (bg_name, bg_indexes, use_simulator_bgs, use_simulator_masks, 0, gt_movement, m_movement)), where:

  • x: masked sequence.

  • m: binary shape used to mask y.

  • y: original background sequence before being masked.

  • bg_name: name of the sequence used in the original dataset.

  • bg_indexes: string

  • use_simulator_bgs: boolean indicating if the background sequence has been obtained by applying random transformations to a frame.

  • use_simulator_masks: boolean indicating if the masks

    sequence has been obtained by applying random transformations to a mask.

  • gt_movement: tensor containing

__len__() → int

Returns the length of the dataset. If the argument frames_n is set to -1, the length is the number of background videos available. If set to another value, the length is the sum of all the sequences frames.

Returns

Number of dataset items using the given configuration.

class master_thesis.dataset.MasterThesisContentProvider(dataset_meta, bgs_movement_min_height, movement_simulator, **kwargs)

Bases: torch.utils.data.dataset.Dataset

__getitem__(frame_index)

Returns both the background and the mask associated with index frame_item.

Parameters

frame_index – frame index between 0 and self.__len__().

Returns

tensor of size (C,H,W) containing a GT image quantized between [0,1] or None. torch.FloatTensor: tensor of size (C,H,W) containing a mask image quantized between [0,1] or None. str: name of the sequence.

Return type

torch.FloatTensor

get_items(frames_indexes)

Returns the backgrounds and the masks at indexes frames_indexes.

Parameters

  • frames_indexes – list of indexes of the frames that should be

  • returned. –

Returns

Tuple of two tensors containing the backgrounds and the masks, respectively. The tensor containing the backgrounds has size (C,F, H,W), while the tensor containing the masks is (1,F,H,W).

__len__()

Returns the sum of the frames of all sequences.

Returns

Length of the dataset, given by the sum of the frames of all sequences.

get_sequence(sequence_index)

Returns the sequence with index sequence_index.

Parameters

  • sequence_index – sequence index between 0 and

  • self.len_sequences() –

Returns

Tuple containing the backgrounds, masks, names

len_sequences()

Return the number of different sequences.

get_patch(frame_index, frames_n, use_movement_simulator)
get_patch_random(frames_n, movement_simulator)