【论文系列研读】Superpixel： SLIC+SNN

【论文系列研读】Superpixel： SLIC+SNN

1、SLIC（PAMI2012）

Title：SLIC Superpixels Compared to State-of-the-art Superpixel Methods

Author：Radhakrishna Achanta ... (École Polytechnique Fédérale de Lausanne,EPFL 瑞士联邦理工学院)

Other Algorithms for generating superpixels

1.Graph-based algorithms

• treat each pixel as a node
• Edge weights are similarity between neighboring pixels.
• bipartite graph
• finding optimal paths

2.Gradient-ascent-based algorithms

算法:

Advantages

• Fastest
• most memory efficient

结果

1. 自然图像

2. 2D and 3D EM images

2、Superpixel Sampling Networks(ECCV2018)

Title：Superpixel Sampling Networks

Author：Varun Jampani ... (NVIDIA)

Why is SLIC not differentiable?

• a non-differentiable nearest neighbor operation
• Associate each pixel to the nearest superpixel center

Advantages：

soft-associations

1. the first end-to-end trainable superpixel algorithm
2. convert the nearest-neighbor operation into differentiable
3. learning with flexible loss functions

算法

• m：superpixel个数
• QF=weighted sum of pixel features，距离为权值，对特征加权
• Optional：求每个superpixel内的最大距离值，最小化这个值
• column normalized Qt as Qˆt

Loss function：

           segmentation tasks： cross-entropy loss

           optical flow ： L1-norm

           compactness loss ：lower spatial variance

结果：