The implemented networks can be found in models/networks/segmentation/:
- Task1: Analyse Datasets: Camvid, CityScapes, Kitti
- Task1: Train FCN8 with Camvid
- Task2: Read papers FCN8 and SEGNET
- Task3: Train SegNet with Camvid: using pretrained VGG16
- Task4: Train FCN8 and SegNet for: Synthis-CityScapes and Kitti
- Task5: Boost performance in both Networks
- Task6: Report and Slides
The following table shows the results obtained from the different tasks. We have used two networks, the Fully Convolutional Network FCN8 and our implementation of the SegNet. The experiments have been done using the following datasets: CAamvid, Synthia-CityScapes and KITTI.
| Network | Experiment | Camvid | KITTI | Synthia | |||
|---|---|---|---|---|---|---|---|
| Val | Test | Val | Test | Val | Test | ||
| FCN8 | Accuracy | 76.25 | 67.33 | 31.03 | - | 70.05 | 70.78 |
| FCN8 | mIoU | 65.25 | 56.81 | 26.92 | - | 63.45 | 64.22 |
| SEGNET | Accuracy | 77.95 | 58.52 | 27.31 | - | 55.30 | 55.37 |
| SEGNET | mIoU | 65.50 | 46.60 | 22.83 | - | 50.00 | 49.75 |