『笔记』DeepFusion

『笔记』DeepFusion

一个point-level fusion的方式，实际上是voxel级别上paint. 虽然用了cross attention不过其关注点其实是在于alignment quality上

3. DeepFusion

3.1. Deep Feature Fusion Pipeline

Motivation: the voxelization module is not designed for processing camera information. Second, the camera feature extractor is learned from other independent tasks (i.e., 2D detection or segmentation), which may lead to (1) domain gap, (2) annotation efforts, (3) additional computation cost, and more importantly, (4) sub-optimal extracted features because the features are heuristically cho- sen rather than learned in an end-to-end manner.

3.2. Impact of Alignment Quality

3.3. Boosting Alignment Quality