PLEASE
A variant of the channel attention module that learns soft region assignments end to end, using a lightweight assignment network to output region membership weights for each pixel.
Timeline (last 90 days)
The paper proposes four region-aware channel recalibration modules—BASE, SPASE, MAPSE and PLEASE—which define pooling regions using regular grids, SLIC superpixels, frozen network semantic maps and end-to-end learned soft assignments, respectively.
PLEASE learns soft region assignments end-to-end through a lightweight assignment network, obtaining a gate value for each channel in each region, bridging channel attention and spatial attention.
PLEASE produces soft assignment results along the way during forward propagation, providing structured spatial interpretability without post-hoc analysis.
The inference computation latency added by PLEASE is below the measurable range.
Experimental results show that the four modules surpass the original SE module and related spatial-channel attention variants on most benchmarks, achieving the best results in the majority of cases.
All Facts (5)
The paper proposes four region-aware channel recalibration modules—BASE, SPASE, MAPSE and PLEASE—which define pooling regions using regular grids, SLIC superpixels, frozen network semantic maps and end-to-end learned soft assignments, respectively.
50%UnverifiedPLEASE learns soft region assignments end-to-end through a lightweight assignment network, obtaining a gate value for each channel in each region, bridging channel attention and spatial attention.
50%UnverifiedPLEASE produces soft assignment results along the way during forward propagation, providing structured spatial interpretability without post-hoc analysis.
50%UnverifiedThe inference computation latency added by PLEASE is below the measurable range.
50%UnverifiedExperimental results show that the four modules surpass the original SE module and related spatial-channel attention variants on most benchmarks, achieving the best results in the majority of cases.
50%