search

dvl-tum / SUSHI

Official PyTorch implementation of Unifying Short and Long-Term Tracking with Graph Hierarchies (CVPR 2023)
MIT License
107 stars 9 forks source link

Weight sharing between levels #14

Closed WesLee88524 closed 1 year ago

WesLee88524 commented 1 year ago

Hi Orcun, I have a question about weight sharing between levels. You said "SUSHI blocks use the same GNN architecture at every hierarchy level. We observe that we can share parameters and learnable weights among the GNNs used for each SUSHI block." in paper. But why train the different layers independent and use different params in code ?

ocetintas commented 1 year ago

We indeed share weights between levels. Please check https://github.com/dvl-tum/SUSHI/blob/main/src/models/hiclnet.py#L21. We use all_but_first setting because 'motion' features are not available for the very first level.