Fig. 1. Architecture of ULNN
Fig. 2. (a) Input data construction procedure of ULNN-rep; (b) input data construction procedure of ULNN-aug
Fig. 3. Components of an UL layer
Fig. 4. Curves of uncertainty learning with different βs (on CIFAR-10, vertical-axis is a logarithmic coordinate)
Fig. 5. Curves of UL (CIFAR-10)
| CIFAR-10 | CIFAR-100 |
---|
| Accuracy | Uncertainty | Accuracy | Uncertainty | DenseNet-rep | 91.6% | 0.0356 | 68.3% | 0.1166 | ULNN-rep | 92.4% | 0.0010 | 69.3% | 0.0066 | ULNN-aug | 94.3% | 0.0828 | 74.2% | 0.1480 |
|
Table 1. Comparison between ULNN-rep and ULNN-aug
nu | CIFAR-10 | CIFAR-100 |
---|
| Accuracy | Uncertainty | Accuracy | Uncertainty | 2 | 94.8% | 0.0579 | 75.9% | 0.0753 | 3 | 94.9% | 0.0659 | 75.6% | 0.0762 | 4 | 95.0% | 0.0717 | 75.7% | 0.0743 | 5 | 94.8% | 0.0761 | 75.6% | 0.0740 |
|
Table 2. Impact of input repetition number nu
β | CIFAR-10 | CIFAR-100 |
---|
1.5(=nu/2) | 0.0718 | 0.1076 | 1 | 0.0828 | 0.1480 | 0.1 | 0.2107 | 0.8285 | 0.01 | 0.6936 | 2.4774 |
|
Table 3. Impact of UL weight β
Dropout ratio | CIFAR-10 | CIFAR-100 |
---|
| Accuracy | Uncertainty | Accuracy | Uncertainty | 0.2 | 94.8% | 0.0757 | 75.2% | 0.1417 | 0.1 | 94.9% | 0.0687 | 75.6% | 0.1077 | 0.05 | 95.1% | 0.0579 | 75.9% | 0.0753 |
|
Table 4. Impact of Dropout on ULNN-aug
| CIFAR-10 | CIFAR-100 |
---|
| Accuracy | Uncertainty | Accuracy | Uncertainty | DenseNet + data aug. | 94.8% | 1.5190 | 75.6% | 1.3564 | ULNN-aug | 95.1% | 0.0579 | 75.9% | 0.0753 |
|
Table 5. Comparison of validation result with original training parameters in DenseNet
| Accuracy | mIoU | Uncertainty |
---|
Bayesian SegNet | 85.9% | 51.4% | 452.7 | ULNN SegNet | 87.1% | 52.0% | 50.63 |
|
Table 6. Semantic segmentation results of ULNN SegNet on CamVid