A three-stage CNN compression framework (dynamic cropping + compound shrinking + cascaded dynamic inference) reduces ResNet-50 MACs by 48.8% while improving ImageNet-1K top-1 accuracy by 0.8%.
Title resolution pending
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
-
EdgeCompress: Coupling Multidimensional Model Compression and Dynamic Inference for EdgeAI
A three-stage CNN compression framework (dynamic cropping + compound shrinking + cascaded dynamic inference) reduces ResNet-50 MACs by 48.8% while improving ImageNet-1K top-1 accuracy by 0.8%.