Method Details
Details for method 'L2-SP'
Method overview
| name | L2-SP |
| challenge | pixel-level semantic labeling |
| details | With a simple variant of weight decay, L2-SP regularization (see the paper for details), we reproduced PSPNet based on the original ResNet-101 using "train_fine + val_fine + train_extra" set (2975 + 500 + 20000 images), with a small batch size 8. The sync batch normalization layer is implemented in Tensorflow (see the code). |
| publication | Explicit Inductive Bias for Transfer Learning with Convolutional Networks Xuhong Li, Yves Grandvalet, Franck Davoine ICML-2018 https://arxiv.org/abs/1802.01483 |
| project page / code | https://github.com/holyseven/PSPNet-TF-Reproduce |
| used Cityscapes data | fine annotations, coarse annotations |
| used external data | ImageNet |
| runtime | n/a |
| subsampling | no |
| submission date | July, 2018 |
| previous submissions | 1 |
Average results
| Metric | Value |
|---|---|
| IoU Classes | 81.1935 |
| iIoU Classes | 58.1385 |
| IoU Categories | 91.0223 |
| iIoU Categories | 78.4897 |
Class results
| Class | IoU | iIoU |
|---|---|---|
| road | 98.6691 | - |
| sidewalk | 86.83 | - |
| building | 93.5765 | - |
| wall | 64.7495 | - |
| fence | 63.3864 | - |
| pole | 67.3632 | - |
| traffic light | 74.4991 | - |
| traffic sign | 79.3964 | - |
| vegetation | 93.6357 | - |
| terrain | 73.0739 | - |
| sky | 95.583 | - |
| person | 86.548 | 67.9365 |
| rider | 72.4557 | 49.8065 |
| car | 96.0909 | 90.1152 |
| truck | 76.0223 | 42.5998 |
| bus | 90.7018 | 52.7551 |
| train | 83.1075 | 51.9492 |
| motorcycle | 70.4696 | 47.6876 |
| bicycle | 76.518 | 62.258 |
Category results
| Category | IoU | iIoU |
|---|---|---|
| flat | 98.7168 | - |
| nature | 93.3593 | - |
| object | 73.6352 | - |
| sky | 95.583 | - |
| construction | 93.6972 | - |
| human | 86.5825 | 68.9244 |
| vehicle | 95.5822 | 88.055 |