Open usersan opened 4 years ago
Decoder | Encoder | Coarse | mIoU | Road | Sidewalk | Building | Sign | Sky | Person | Car | Bicycle | Truck |
---|---|---|---|---|---|---|---|---|---|---|---|---|
SkipNet | MobileNet | No | 61.3 | 95.9 | 73.6 | 86.9 | 57.6 | 91.2 | 66.4 | 89.0 | 63.6 | 45.9 |
SkipNet | ShuffleNet | No | 55.5 | 94.8 | 68.6 | 83.9 | 50.5 | 88.6 | 60.8 | 86.5 | 58.8 | 29.6 |
UNet | ResNet18 | No | 57.9 | 95.8 | 73.2 | 85.8 | 57.5 | 91.0 | 66.0 | 88.6 | 63.2 | 31.4 |
UNet | MobileNet | No | 61.0 | 95.2 | 71.3 | 86.8 | 60.9 | 92.8 | 68.1 | 88.8 | 65.0 | 41.3 |
UNet | ShuffleNet | No | 57.0 | 95.1 | 69.5 | 83.7 | 54.3 | 89.0 | 61.7 | 87.8 | 59.9 | 35.5 |
Dilation | MobileNet | No | 57.8 | 95.6 | 72.3 | 85.9 | 57.0 | 91.4 | 64.9 | 87.8 | 62.8 | 26.3 |
Dilation | ShuffleNet | No | 53.9 | 95.2 | 68.5 | 84.1 | 57.3 | 90.3 | 62.9 | 86.6 | 60.2 | 23.3 |
SkipNet | MobileNet | Yes | 62.4 | 95.4 | 73.9 | 86.6 | 57.4 | 91.1 | 65.7 | 88.4 | 63.3 | 45.3 |
SkipNet | ShuffleNet | Yes | 59.3 | 94.6 | 70.5 | 85.5 | 54.9 | 90.8 | 60.2 | 87.5 | 58.8 | 45.4 |
Cityscapesで評価。解像度512x1024。 デコーダ精度:UNet >= SkipNet > Dilation
Model | GFLOPs | FPS |
---|---|---|
SkipNet-MobileNet | 13.8 | 45 |
UNet-MobileNet | 55.7 | 20 |
UNetよりSkipNetの方が軽い
Model | GFLOPs | Class IoU | Class iIoU | Category IoU | Category iIoU |
---|---|---|---|---|---|
SegNet[21] | 286.03 | 56.1 | 34.2 | 79.8 | 66.4 |
ENet[12] | 3.83 | 58.3 | 24.4 | 80.4 | 64.0 |
DeepLab[2] | - | 70.4 | 42.6 | 86.4 | 67.7 |
SkipNet-VGG16[1] | - | 65.3 | 41.7 | 85.7 | 70.1 |
SkipNet-ShuffleNet | 2.0 | 58.3 | 32.4 | 80.2 | 62.2 |
SkipNet-MobileNet | 6.2 | 61.5 | 35.2 | 82.0 | 63.0 |
解像度360x640。 DeepLabは性能は高いが計算効率は悪い。 計算効率の面では、ENet、SkipNet-ShuffleNetやSkipNet-MobileNetでの比較になるか。 SkipNet-ShuffleNetはENetより軽い。 エンコーダはMobileNet、ShuffleNet、ResNet18だとMobileNetが正確。
0. 論文
Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand
https://arxiv.org/abs/1803.02758
1. どんなもの?
複数の特徴抽出器(VGG16、Resnet18、MobileNet、ShuffleNet)とデコーダのアーキテクチャ(SkipNet、UNet、Dilation Frontend)を組み合わせて精度・計算量を比較
2. 先行研究と比べてどこがすごい?
3. 技術や手法のキモはどこ?
4. どうやって有効だと検証した?
5. 議論はある?
6. 次に読むべき論文は?