DICTA 2019
Juntao Li^1
College of Engineering and Computer Science, The Australian National University
Chuong Nguyen^2, Shaodi You
DATA61 - Cyber Physical Systems, CSIRO, Canberra, Australia
The reporsitory is still under construction :)
The figure shows one example of water detection where red areas are water areas.
The figure shows the network structure of our paper, our proposed network introduces temporal dimensions into FCN using 3D convolutional neural network which gives compariable results as [1] while reduce the time and space consumption significantly.
POOLING COMPARISON ON ON-ROAD DATASET
| Pool | F1-measure | Precision | Recall | Time(sec/frame) |
|---|---|---|---|---|
| Average(ours) | 0.648 | 0.774 | 0.578 | 0.23103 |
| Max(ours) | 0.680 | 0.710 | 0.662 | 0.23001 |
| Last(ours) | 0.684 | 0.788 | 0.616 | 0.22975 |
PERFORMANCE COMPARISON ON ON-ROAD DATASET
| Method | F1-measure | Precision | Recall | Time(sec/frame) |
|---|---|---|---|---|
| T3D-FCN-LAST(ours) | 0.68 | 0.79 | 0.62 | 0.23 |
| 2D-FCN (ours) | 0.51 | 0.51 | 0.49 | 0.20 |
| FCN-8s-FL-RAU[8] | 0.70 | 0.68 | 0.72 | 0.32 |
| FCN-8s[14] | 0.57 | 0.59 | 0.55 | 0.06 |
| DeepLab[4] | 0.22 | 0.37 | 0.16 | 0.06 |
| GMM & polar[15] | 0.31 | 0.19 | 0.90 | NA |