search

YihongSun / Bayesian-Amodal

[CVPR 2022] Amodal Segmentation through Out-of-Task and Out-of-Distribution Generalization with a Bayesian Model
MIT License
22 stars 4 forks source link

How is this method implemented in the code? #7

Closed FedericaGielen closed 1 year ago

FedericaGielen commented 1 year ago

Are eight shape templates devised for each object class？How is shape modeling implemented in the code? Would increasing the number of shape template of each object class improve the segmentation accuracy? 25 68

YihongSun commented 1 year ago

Are eight shape templates devised for each object class？

Yes. The templates are constructed for each object class. Please refer to the supplementary materials for more visualizations.

How is shape modeling implemented in the code?

The implementation is in Code/Learn_mix_model_view_CA_it.py. Please refer to L267-L306 for the construction of the shape prior. In essence, we apply a rough foreground-background binary segmentation in the feature space (L56) and compile the foreground and context probabilities across all examples within the cluster

Would increasing the number of shape template of each object class improve the segmentation accuracy?

Yes, it would be an efficiency-performance trade-off since adding more templates would improve disentanglement of varying poses, but also increase the search space at inference time