notesum.ai
Published at October 30Decoupling Semantic Similarity from Spatial Alignment for Neural Networks
cs.CV
cs.AI
cs.LG
Released Date: October 30, 2024
Authors: Tassilo Wald1, Constantin Ulrich1, Gregor Köhler1, David Zimmerer1, Stefan Denner1, Michael Baumgartner1, Fabian Isensee1, Priyank Jaini2, Klaus H. Maier-Hein1
Aff.: 1German Cancer Research Center (DKFZ), Heidelberg, Germany; 2Google Deepmind

| Pearson Correlation | ||||||
|---|---|---|---|---|---|---|
| Architectures | Cosine Sim. | Inner Product | RBF | |||
| - | (ours) PI | - | (ours) PI | - | (ours) PI | |
| ResNet18 | -0.276 | -0.326 | -0.259 | -0.270 | -0.176 | -0.199 |
| ResNet50 | -0.248 | -0.291 | -0.243 | -0.261 | 0.040 | 0.029 |
| ResNet101 | -0.192 | -0.276 | -0.174 | -0.240 | 0.091 | 0.084 |
| ConvNextV2-Base | -0.134 | -0.098 | -0.132 | -0.171 | 0.117 | 0.090 |
| ViT-B/16 | -0.046 | -0.100 | -0.045 | -0.026 | -0.077 | -0.122 |
| ViT-L/32 | -0.138 | -0.188 | -0.138 | -0.144 | -0.134 | -0.166 |
| DinoV2-Giant | -0.012 | -0.044 | -0.013 | -0.031 | -0.008 | -0.048 |