Video transforms

Eztorch supports transforms from Torchaug and from Pytorchvideo.

On top of that, several video transforms have been defined.

Video transforms

class eztorch.transforms.video.RemoveTimeDim[source]

Remove time dimension from tensor.

Suppose the tensor shape is [C,T,H,W].

class eztorch.transforms.video.RandomResizedCrop(target_height, target_width, scale, aspect_ratio, shift=False, log_uniform_ratio=True, interpolation='bilinear', num_tries=10)[source]

nn.Module wrapper for pytorchvideo.transforms.functional.random_resized_crop.

class eztorch.transforms.video.RandomTemporalDifference(use_grayscale=True, absolute=False, p=0.2)[source]

Randomly compute temporal differences from consecutives frames.

Parameters:
  • use_grayscale (bool, optional) – If True, apply grayscale transformation before computing the temporal difference.

    Default: True

  • absolute (bool, optional) – If True, take the absolute value of the difference, else shift to the mean value (\(255\) for int, \(0.5\) for float) and divide by 2.

    Default: False

  • p (float, optional) – The probability to compute temporal difference.

    Default: 0.2

class eztorch.transforms.video.TemporalDifference(use_grayscale=True, absolute=False)[source]

Compute temporal differences from consecutives frames.

Parameters:
  • use_grayscale (bool, optional) – If True, apply grayscale transformation before computing the temporal difference.

    Default: True

  • absolute (bool, optional) – If True, take the absolute value of the difference, else shift to the mean value (\(255\) for int, \(0.5\) for float) and divide by 2.

    Default: False

SoccerNet

class eztorch.transforms.video.soccernet.BatchReduceTimestamps[source]

Aggregate successive timestamps of soccernet batch of clips information.

class eztorch.transforms.video.soccernet.BatchMiddleTimestamps[source]

Select the middle timestamps of soccernet batch of clips information.

class eztorch.transforms.video.soccernet.ReduceTimestamps[source]

Aggregate successive timestamps of soccernet clip information.

class eztorch.transforms.video.soccernet.MiddleTimestamps[source]

Select the middle timestamps of soccernet clip information.

Spot

class eztorch.transforms.video.spot.BatchMiddleFrames[source]

Select the middle frame of spot batch of clips information.

class eztorch.transforms.video.spot.MiddleFrames[source]

Select the middle frame of spot clip information.

class eztorch.transforms.video.spot.SpottingMixup(alpha=0.5)[source]

Make mixup for spotting for labels.

Parameters:

alpha (float, optional) – Alpha value for the beta distribution of mixup.

Default: 0.5