Fully Convolutional Networks for Semantic Segmentation
Jonathan Long, Evan Shelhamer, Trevor Darrell. "Fully Convolutional Networks for Semantic Segmentation." 2015.
Models
Using Dataset
| Implementation | Accuracy | Weights | Memory | Conv Ops | etc | link |
|---|---|---|---|---|---|---|
| Keras | FCN8: 427,127,544, FCN32 : 4,234,357,544 | Keras | ||||
| Tensorflow Slim | 256,445,037 | 256,445,037 * 4bytes | Slim | |||
| Pytorch | 85.62% (Pixel-Acc) | 134,489,759 | 134,489,759 * 4bytes | W.I.P (Voc2012) | Pytorch |
Tip & Trick
| name | for What | reference |
|---|---|---|
| Skip Layers | to produce accurate and detailed segmentations | - |
| Becoming Fully Convolutional | 1. to be free from the size of image 2. to store location information of each pixel |
- |
| Dilated ConV | 1. Dilated convolution for extract global feature | https://arxiv.org/abs/1511.07122 |
Error of paper
- add this if any errors