vis4d.model.detect3d.bevformer

BEVFromer model implementation.

This file composes the operations associated with BEVFormer https://arxiv.org/abs/2203.17270 into the full model implementation.

Classes

BEVFormer(basemodel[, fpn, pts_bbox_head, ...])

BEVFormer 3D Detector.

PrevFrameInfo

Previous frame information.
class BEVFormer(basemodel, fpn=None, pts_bbox_head=None, weights=None)[source]

BEVFormer 3D Detector.

Creates an instance of the class.

Parameters:

  • basemodel (BaseModel) – Base model network.

  • fpn (FPN, optional) – Feature Pyramid Network. Defaults to None. If None, a default FPN will be used.

  • pts_bbox_head (BEVFormerHead, optional) – BEVFormer head. Defaults to None. If None, a default BEVFormer head will be used.

  • weights (str, optional) – Path to the checkpoint to load. Defaults to None.

extract_feat(images_list)[source]

Extract features of images.

Return type:

list[Tensor]

forward(images, can_bus, scene_names, cam_intrinsics, cam_extrinsics, lidar_extrinsics)[source]

Forward.

Return type:

Detect3DOut

class PrevFrameInfo[source]

Previous frame information.