Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials (2011)
- Traditional Vision Techniques
- Not Deep Learning
- Man-designed Feature extraction
- Contribution Points
- Use Dense CRF for representation of relationship between pixel
- Preserve color channel difference
- == Preserve Edges
- Induce bilateral filter based on gaussian filter
- Formulate Meanfield inference for fast computation
- Use Permutohedral Lattice for computing gaussian filter
- Offical Code and slides
Fast High-Dimensional Filtering Using the Permutohedral Lattice (2010)
- Contribution Points
- reduce time complexity of $O(e^d)$ into $O(n^2n)$ for High-Dimensional Gaussian Filtering
- e.g. Bilateral Filter: 2D pixel location + 3D color channel = 5D position info(input for individual pixel, not location)
- $ \vec{v_i} = \sum\limits^{n}_{j}e^{-\frac{1}{2}|p_i - p_j|^2}\vec{v_j}$
- Permutohedral Lattice
- Projection D dimension High Dimensional Positions()
Fully Convolutional Networks for Semantic Segmentation (2015)
- Contribution Points
- The first paper deeling with semantic segmentation using FCN (Fully Convolutional Networks)
- Code
Many papers using Full Deconv Net symmetric as Front CNNs (2015)
SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
Semantic Image Segmentation with Deep Convlutional Nets and Fully Connected CRFs (2015)
- Contribution Points
- Use CRFs as Post Processing
- Code
Conditional Random Fields as Recurrent Neural Networks (2015)
- Contribution Points
- Implement CRF as Recurrent CNN
- Code
Semantic Segmentation using Adversarial Network
- Set Segmentation Networks as Typical G(z)
- invent new network D(x) for matching function, If ground truth class map and segmentation map from G(z) is similar -> 1 otherwise 0
- New D(x)
- Input: Target image to segment and Class map
- ground truth or Segmented map from G(z) Here z is the target image, There is no noise Z
Suggestions Network from comma.ai (SugNet)