DVP+ extends a single-object disentangling autoencoder to multi-object scenes, using alternative training losses and per-scene optimization to outperform MONet and LIVE on a new benchmark, when the object count is known.
Matched candidates serve as the positive class, unmatched candidates as the negative class, and Binary Cross-Entropy (BCE) is used for this loss (Eq
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Generative Learning of Differentiable Object Models for Compositional Interpretation of Complex Scenes
DVP+ extends a single-object disentangling autoencoder to multi-object scenes, using alternative training losses and per-scene optimization to outperform MONet and LIVE on a new benchmark, when the object count is known.