Embedding is all your need.
- The embedding could be easily exponental averaged between frames for better tackling with occlusion.
- Lidar perception can also benefited from this architecture which make
the appresiating
mid-fusionscheme easy and efficient.
- Unify the embedding expression between image and texture to ease and improve the multi-model works such as image caption and image retrial and multi-model comprehension.
- Learning to extract an additional positional free embedding which we call it semantic/texture/apperance embedding.
- Make the scene/thing and shape/appearence lantent embedding code more meaningful
and releaf the buddern of the network expression and learning for
GIRAFFE.
- Faster than
RealTimePanon Cityscapes and Mapli Dataset with network architecture adaptation. - More accurate than current
PanopticFCNby unify the training and inference embedding expression with affilant loss. - More accurate than current
PanopticFCNby making the embedding contain more accurate positional information with affilant loss.