vis4d.op.seg.fcn¶

FCN Head for semantic segmentation.

Classes

`FCNHead`(in_channels, out_channels[, ...])	FCN Head made with ResNet base model.
`FCNOut`(pred, outputs)	Output of the FCN prediction.

class FCNHead(in_channels, out_channels, dropout_prob=0.1, resize=None)[source]¶

FCN Head made with ResNet base model.

This is based on the implementation in torchvision.

Creates an instance of the class.

Parameters:

in_channels (list[int]) – Number of channels in multi-level image feature.
out_channels (int) – Number of output channels. Usually the number of classes.
dropout_prob (float, optional) – Dropout probability. Defaults to 0.1.
resize (tuple(int,int), optional) – Target shape to resize output. Defaults to None.

__call__(feats)[source]¶

Type definition for function call.

forward(feats)[source]¶

Transforms feature maps and returns segmentation prediction.

Parameters:

feats (list[torch.Tensor]) – List of multi-level image features.

Returns:

Each tensor has shape (batch_size, self.channels, H, W) which is prediction for each FCN stages. E.g.,

outputs[-1] ==> main output map outputs[-2] ==> aux output map (e.g., used for training) outputs[:-2] ==> x[:-2]

Return type:

output (list[torch.Tensor])

class FCNOut(pred: torch.Tensor, outputs: list[torch.Tensor])[source]¶

Output of the FCN prediction.

Create new instance of FCNOut(pred, outputs)