vis4d.model.detect.faster_rcnn

Faster RCNN model implementation and runtime.

Classes

FasterRCNN(num_classes[, basemodel, ...])

Faster RCNN model.
class FasterRCNN(num_classes, basemodel=None, faster_rcnn_head=None, rcnn_box_decoder=None, weights=None)[source]

Faster RCNN model.

Creates an instance of the class.

Parameters:

  • num_classes (int) – Number of object categories.

  • basemodel (BaseModel, optional) – Base model network. Defaults to None. If None, will use ResNet50.

  • faster_rcnn_head (FasterRCNNHead, optional) – Faster RCNN head. Defaults to None. if None, will use default FasterRCNNHead.

  • rcnn_box_decoder (DeltaXYWHBBoxDecoder, optional) – Decoder for RCNN bounding boxes. Defaults to None.

  • weights (str, optional) – Weights to load for model. If set to “mmdet”, will load MMDetection pre-trained weights. Defaults to None.

__call__(images, input_hw, boxes2d=None, boxes2d_classes=None, original_hw=None)[source]

Type definition for call implementation.

Return type:

FRCNNOut | DetOut

forward(images, input_hw, boxes2d=None, boxes2d_classes=None, original_hw=None)[source]

Forward pass.

Parameters:

  • images (torch.Tensor) – Input images.

  • input_hw (list[tuple[int, int]]) – Input image resolutions.

  • boxes2d (None | list[torch.Tensor], optional) – Bounding box labels. Required for training. Defaults to None.

  • boxes2d_classes (None | list[torch.Tensor], optional) – Class labels. Required for training. Defaults to None.

  • original_hw (None | list[tuple[int, int]], optional) – Original image resolutions (before padding and resizing). Required for testing. Defaults to None.

Returns:

Either raw model outputs (for training) or

predicted outputs (for testing).

Return type:

FRCNNOut | DetOut

forward_test(images, images_hw, original_hw)[source]

Forward testing stage.

Parameters:

  • images (torch.Tensor) – Input images.

  • images_hw (list[tuple[int, int]]) – Input image resolutions.

  • original_hw (list[tuple[int, int]]) – Original image resolutions (before padding and resizing).

Returns:

Predicted outputs.

Return type:

DetOut

forward_train(images, images_hw, target_boxes, target_classes)[source]

Forward training stage.

Parameters:

  • images (torch.Tensor) – Input images.

  • images_hw (list[tuple[int, int]]) – Input image resolutions.

  • target_boxes (list[torch.Tensor]) – Bounding box labels.

  • target_classes (list[torch.Tensor]) – Class labels.

Returns:

Raw model outputs.

Return type:

FRCNNOut