Introduction
Related work
Unpaired translation
Contrastive learning
Methods
Adversarial loss
PatchNCE loss
MCL loss
Final objective loss
Experiments
Training details
Datasets
Evaluation protocol
Method | Horse\(\rightarrow \)Zebra | Cat\(\rightarrow \)Dog | CityScapes | |||||
---|---|---|---|---|---|---|---|---|
sec/iter\(\downarrow \) | Model parameters\(\downarrow \) | FID | FID | FID | pixAcc\(\downarrow \) | classAcc\(\downarrow \) | IoU\(\downarrow \) | |
CycleGAN [47] | 0.40 | 28.286M | 77.2 | 85.9 | 76.3 | 0.52 | 0.17 | 0.11 |
GcGAN [11] | 0.26 | 16.908M | 86.7 | 96.6 | 105.2 | 0.55 | 0.20 | 0.13 |
FastCUT [36] | 0.15 | 14.703M | 73.4 | 94.0 | 68.8 | 0.65 | 0.21 | 0.15 |
CUT [36] | 0.24 | 14.703M | 45.5 | 76.2 | 56.4 | 0.70 | 0.24 | 0.17 |
SimDCL [16] | 0.47 | 28.852M | 47.1 | 65.5 | 51.3 | 0.69 | 0.21 | 0.15 |
DCLGAN [16] | 0.41 | 28.812M | 43.2 | 60.7 | 49.4 | 0.74 | 0.22 | 0.17 |
FastMCL(ous) | 0.15 | 14.703M | 46.5 | 88.8 | 55.3 | 0.76 | 0.25 | 0.19 |
MCL(ous) | 0.25 | 14.703M | 40.7 | 70.2 | 47.3 | 0.78 | 0.26 | 0.21 |
Unpaired image translation
Single image translation
Ablation study
Method | Min | Max | Mean | SD |
---|---|---|---|---|
FastCUT [36] | 40.8 | 116.9 | 75.6 | 21.7 |
CUT [36] | 44.0 | 78.2 | 56.8 | 10.7 |
MCL(ous) | 41.4 | 68.8 | 51.9 | 7.6 |
Model | \({{\lambda }_{X}}\) | \({{\lambda }_{Y}}\) | \({{\lambda }_{M}}\) | FID |
---|---|---|---|---|
FastCUT [36] | 1 | \(\times \) | \(\times \) | 73.4 |
FastMCL | 1 | \(\times \) | 0.1 | 70.0 |
FastMCL (ours) | 1 | \(\times \) | 0.01 | 46.5 |
CUT [36] | 1 | 1 | \(\times \) | 45.5 |
MCL | 1 | 1 | 0.1 | 43.9 |
MCL(ours) | 1 | 1 | 0.01 | 40.7 |
Method | sec/iter\(\downarrow \) | Model Parameters\(\downarrow \) | FID |
---|---|---|---|
CycleGAN [47] | 0.40 | 28.286M | 77.2 |
CycleGAN + MCL loss | 0.41 | 28.286M |
70.1
|
DCLGAN [16] | 0.41 | 28.812M | 43.2 |
DCLGAN + MCL loss | 0.42 | 28.812M |
39.6
|
SimDCL [16] | 0.47 | 28.852M | 47.1 |
SimDCL + MCL loss | 0.48 | 28.852M |
39.7
|