Towards Learning Monocular 3D Object Localization From 2D Labels using the Physical Laws of Motion

Segformer++: Efficient Token-Merging Strategies for High-Resolution Semantic Segmentation