A fixed concept-to-part routing plus a Gaussian prior from average keypoint positions grounds CBM attention at ~89% top-1 and ~70% pointing accuracy with no per-image spatial labels.
arXiv preprint arXiv:2305.19550 , year =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Spatially Grounded Concept Bottleneck Models via Part-Factorized Attention
A fixed concept-to-part routing plus a Gaussian prior from average keypoint positions grounds CBM attention at ~89% top-1 and ~70% pointing accuracy with no per-image spatial labels.