Multi-object semantic discovery on Multi-dSprites
91.3mAccDTI-Sprites
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DTI-SpritesSprite generation=Pixels, Decision function=Min-Loss, Activation function=S, Training criterion=L0−1=comp, Regularization=Lempty, Training schedule=Longer training (†)2026.04 | 91.3 | 84 | |
| Ours-DSprite generation=MLP, Decision function=linear mapping, Activation function=GS, Training criterion=Lcomp, Regularization=Lempty + Lfreq2026.04 | 66 | 54.7 | |
| Ours-DSprite generation=MLP, Decision function=linear mapping, Activation function=GS, Training criterion=Lcomp, Regularization=Lempty2026.04 | 65.9 | 54.7 | |
| Ours-DSprite generation=MLP, Decision function=linear mapping, Activation function=GS, Training criterion=Lcomp, Regularization=None2026.04 | 65.6 | 54.4 | |
| Ours-DSprite generation=MLP, Decision function=linear mapping, Activation function=GS, Training criterion=Lcomp, Regularization=Lempty + Lbin2026.04 | 65.4 | 54.2 | |
| Ours-DSprite generation=MLP, Decision function=linear mapping, Activation function=GS, Training criterion=Lcomp, Regularization=τ2026.04 | 65.2 | 53.8 |