Method Details
Details for method 'FC-HarDNet-70'
Method overview
| name | FC-HarDNet-70 |
| challenge | pixel-level semantic labeling |
| details | Fully Convolutional Harmonic DenseNet 70 U-shape encoder-decoder structure with HarDNet blocks Trained with single scale loss at stride-4 validation mIoU=77.7 |
| publication | HarDNet: A Low Memory Traffic Network Ping Chao, Chao-Yang Kao, Yu-Shan Ruan, Chien-Hsiang Huang, Youn-Long Lin ICCV 2019 https://arxiv.org/abs/1909.00948 |
| project page / code | https://github.com/PingoLH/FCHarDNet |
| used Cityscapes data | fine annotations |
| used external data | ImageNet |
| runtime | 0.015 s Titan V |
| subsampling | no |
| submission date | October, 2019 |
| previous submissions |
Average results
| Metric | Value |
|---|---|
| IoU Classes | 75.8585 |
| iIoU Classes | 51.3705 |
| IoU Categories | 89.9399 |
| iIoU Categories | 76.6546 |
Class results
| Class | IoU | iIoU |
|---|---|---|
| road | 98.5122 | - |
| sidewalk | 85.4509 | - |
| building | 92.5003 | - |
| wall | 49.0418 | - |
| fence | 54.4441 | - |
| pole | 63.9704 | - |
| traffic light | 71.5329 | - |
| traffic sign | 75.6338 | - |
| vegetation | 93.015 | - |
| terrain | 70.5502 | - |
| sky | 95.3597 | - |
| person | 84.5298 | 65.0336 |
| rider | 67.3695 | 43.0293 |
| car | 95.6614 | 89.1567 |
| truck | 67.7129 | 34.03 |
| bus | 79.0054 | 45.1613 |
| train | 63.6254 | 37.3528 |
| motorcycle | 60.657 | 38.8138 |
| bicycle | 72.7394 | 58.3867 |
Category results
| Category | IoU | iIoU |
|---|---|---|
| flat | 98.6551 | - |
| nature | 92.7812 | - |
| object | 70.4994 | - |
| sky | 95.3597 | - |
| construction | 92.7134 | - |
| human | 84.6337 | 66.3379 |
| vehicle | 94.9367 | 86.9713 |