Embed

Embedding is all your need.

Further work

Video

1. The embedding could be easily exponental averaged between frames for better tackling with occlusion.

Lidar

1. Lidar perception can also benefited from this architecture which make the appresiating mid-fusion scheme easy and efficient.

MultiModel

1. 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.

3D dynamic scene generation

1. Learning to extract an additional positional free embedding which we call it semantic/texture/apperance embedding.
2. Make the scene/thing and shape/appearence lantent embedding code more meaningful and releaf the buddern of the network expression and learning for GIRAFFE.

Panoptic

1. Faster than RealTimePan on Cityscapes and Mapli Dataset with network architecture adaptation.
2. More accurate than current PanopticFCN by unify the training and inference embedding expression with affilant loss.
3. More accurate than current PanopticFCN by making the embedding contain more accurate positional information with affilant loss.