# Implementing Lucas-Kanade Optical Flow algorithm in Python

In this article an implementation of the Lucas-Kanade optical flow algorithm is going to be described. This problem appeared as an assignment in this computer vision course from UCSD. The inputs will be sequences of images (subsequent frames from a video) and the algorithm will output an optical flow field (u, v) and trace the motion of the moving objects. The problem description is taken from the assignment itself.

## Problem Statement

#### Single-Scale Optical Flow

• Let’s implement the single-scale Lucas-Kanade optical flow algorithm. This involves finding the motion (u, v) that minimizes the sum-squared error of the brightness constancy equations for each pixel in a window.  The algorithm will be implemented as a function with the following inputs:

def optical_flow(I1, I2, window_size, tau) # returns (u, v)

• Here, u and v are the x and y components of the optical flow, I1 and I2 are two images taken at times t = 1 and t = 2 respectively, and window_size is a 1 × 2 vector storing the width and height of the window used during flow computation.
• In addition to these inputs, a theshold τ should be added, such that if τ is larger than the smallest eigenvalue of A’A, then the the optical flow at that position should not be computed. Recall that the optical flow is only valid in regions where

has rank 2, which is what the threshold is checking. A typical value for τ is 0.01.

• We should try experimenting with different window sizes and find out the tradeoffs associated with using a small vs. a large window size.
• The following figure describes the algorithm, which considers a nxn (n>=3) window around each pixel and solves a least-square problem to find the best flow vectors for the pixel.

• The following code-snippet shows how the algorithm is implemented in python for a gray-level image.
```import numpy as np
from scipy import signal
def optical_flow(I1g, I2g, window_size, tau=1e-2):

kernel_x = np.array([[-1., 1.], [-1., 1.]])
kernel_y = np.array([[-1., -1.], [1., 1.]])
kernel_t = np.array([[1., 1.], [1., 1.]])#*.25
w = window_size/2 # window_size is odd, all the pixels with offset in between [-w, w] are inside the window
I1g = I1g / 255. # normalize pixels
I2g = I2g / 255. # normalize pixels
# Implement Lucas Kanade
# for each point, calculate I_x, I_y, I_t
mode = 'same'
fx = signal.convolve2d(I1g, kernel_x, boundary='symm', mode=mode)
fy = signal.convolve2d(I1g, kernel_y, boundary='symm', mode=mode)
ft = signal.convolve2d(I2g, kernel_t, boundary='symm', mode=mode) +
signal.convolve2d(I1g, -kernel_t, boundary='symm', mode=mode)
u = np.zeros(I1g.shape)
v = np.zeros(I1g.shape)
# within window window_size * window_size
for i in range(w, I1g.shape[0]-w):
for j in range(w, I1g.shape[1]-w):
Ix = fx[i-w:i+w+1, j-w:j+w+1].flatten()
Iy = fy[i-w:i+w+1, j-w:j+w+1].flatten()
It = ft[i-w:i+w+1, j-w:j+w+1].flatten()
#b = ... # get b here
#A = ... # get A here
# if threshold τ is larger than the smallest eigenvalue of A'A:
nu = ... # get velocity here
u[i,j]=nu[0]
v[i,j]=nu[1]

return (u,v)

```

## Some Results

• The following figures and animations show the results of the algorithm on a few image sequences. Some of these input image sequences / videos are from the course and some are collected from the internet.
• As can be seen, the algorithm performs best if the motion of the moving object(s) in between consecutive frames is slow. To the contrary, if the motion is large, the algorithm fails and we should implement / use multiple-scale version Lucas-Kanade with image pyramids.
• Finally,  with small window size,  the algorithm captures subtle motions but not large motions. With large size it happens the other way.

Input Sequences

Output Optical Flow with different window sizes

window size = 15

window size = 21

Input Sequences

Output Optical Flow

Input Sequences (hamburg taxi)

Output Optical Flow

Input Sequences

Output Optical Flow

Input Sequences

Output Optical Flow

Input Sequences

Output Optical Flow

Input Sequences

Output Optical Flow

Input Sequences

Output Optical Flow

Input Sequences
Output Optical Flow

Input Sequences

Output Optical Flow

Output Optical Flow

Input Sequences

Output Optical Flow with window size 45

Output Optical Flow with window size 10

Output Optical Flow with window size 25

Output Optical Flow with window size 45

# Efficient Graph-Based Image Segmentation in Python

In this article, an implementation of an efficient graph-based image segmentation technique will be described, this algorithm was proposed by Felzenszwalb et. al. from MIT in this paper.  The slides on this paper can be found from this link from the Stanford Vision Lab too. The algorithm is closely related to Kruskal’s algorithm for constructing a minimum spanning tree of a graph, as stated by the author and hence can  be implemented to run in O(m log m) time, where m is the number of edges in the graph.

## Problem Definition and the basic idea (from the paper)

• Let G = (V, E) be an undirected graph with vertices vi ∈ V, the set of elements to be segmented, and edges (vi, vj ) ∈ E corresponding to pairs of neighboring vertices. Each edge (vi, vj ) ∈ E has a corresponding weight w((vi, vj)), which is a non-negative measure of the dissimilarity between neighboring elements vi and vj.

• In the case of image segmentation, the elements in V are pixels and the weight of an edge is some measure of the dissimilarity between the two pixels connected by that edge (e.g., the difference in intensity, color, motion, location or some other local attribute).

• Particularly for the implementation described here, an edge weight function based on the absolute intensity difference (in the yiq space) between the pixels connected by an edge, w((vi, vj )) = |I(pi) − I(pj )|.

• In the graph-based approach, a segmentation S is a partition of V into components
such that each component (or region) C ∈ S corresponds to a connected component
in a graph G0 = (V, E0), where E0 ⊆ E.

• In other words, any segmentation is induced by a subset of the edges in E. There are different ways to measure the quality of a segmentation but in general we want the elements in a component to be similar, and elements in different components to be dissimilar.

• This means that edges between two vertices in the same component should have relatively low weights, and edges between vertices in different components should have higher weights.

• The next figure shows the steps in the algorithm. The algorithm is very similar to Kruskal’s algorithm for computing the MST for an undirected graph.

• The threshold function τ controls the degree to which the difference between two
components must be greater than their internal differences in order for there to be
evidence of a boundary between them.

• For small components, Int(C) is not a good estimate of the local characteristics of the data. In the extreme case, when |C| = 1, Int(C) = 0. Therefore, a threshold function based on the size of the component, τ (C) = k/|C| is needed to be usedwhere |C| denotes the size of C, and k is some constant parameter.

• That is, for small components we require stronger evidence for a boundary. In practice k sets a scale of observation, in that a larger k causes a preference for larger components.

• In general, a Gaussian filter is used to smooth the image slightly before computing the edge weights, in order to compensate for digitization artifacts. We always use a Gaussian with σ = 0.8, which does not produce any visible change to the image but helps remove artifacts.

• The following python code shows how to create the graph.

```import numpy as np
from scipy import signal
import matplotlib.image as mpimg

def gaussian_kernel(k, s = 0.5):
# generate a (2k+1)x(2k+1) gaussian kernel with mean=0 and sigma = s
probs = [exp(-z*z/(2*s*s))/sqrt(2*pi*s*s) for z in range(-k,k+1)]
return np.outer(probs, probs)

def create_graph(imfile, k=1., sigma=0.8, sz=1):
# create the pixel graph with edge weights as dissimilarities
rgb = mpimg.imread(imfile)[:,:,:3]
gauss_kernel = gaussian_kernel(sz, sigma)
for i in range(3):
rgb[:,:,i] = signal.convolve2d(rgb[:,:,i], gauss_kernel, boundary='symm', mode='same')
yuv = rgb2yiq(rgb)
(w, h) = yuv.shape[:2]
edges = {}
for i in range(yuv.shape[0]):
for j in range(yuv.shape[1]):
#compute edge weight for nbd pixel nodes for the node i,j
for i1 in range(i-1, i+2):
for j1 in range(j-1, j+2):
if i1 == i and j1 == j: continue

if i1 >= 0 and i1 = 0 and j1 < h:
wt = np.abs(yuv[i,j,0]-yuv[i1,j1,0])
n1, n2 = ij2id(i,j,w,h), ij2id(i1,j1,w,h)
edges[n1, n2] = edges[n2, n1] = wt
return edges
```

## Some Results

• The images are taken from the paper itself or from the internet. The following figures and animations show the result of segmentation as a result of iterative merging of the components (by choosing least weight edges), depending on the internal difference of the components.

• Although in the paper the author described the best value of the parameter k to be around 300, but  since in this implementation the pixel RGB values are normalized (to have values in between 0 – 1) and then converted to YIQ values and the YIQ intensities are used for computing the weights (which are typically very small), the value of k that works best in this scenario is 0.001-0.01.

• As we can see from the below results, higher the value of the parameter k, larger the size of the final component and lesser the number of components in the result.

• The minimum spanning tree creation is also shown, the red edges shown in the figures are the edges chosen by the algorithm to merge the components.

## Input Image

Output Images for two different values of the parameter k

## Input Image

Output Images for two different values of the parameter k

## Input Image

Output Segmented Images

## Input Image

Output Images for two different values of the parameter k

## Input Image

Segmented Output Image

# Interactive Image Segmentation with Graph-Cut in Python

In this article, interactive image segmentation with graph-cut is going to be discussed. and it will be used to segment the source object from the background in an image. This segmentation technique was proposed by Boycov and Jolli in this paper. This problem appeared as a homework assignment here., and also in this lecture video from the Coursera image processing course by Duke university.

## Problem Statement: Interactive graph-cut segmentation

Let’s implement “intelligent paint” interactive segmentation tool using graph cuts algorithm on a weighted image grid. Our task will be to separate the foreground object from the background in an image.

Since it can be difficult sometimes to automatically define what’s foreground and what’s background for an image, the user is going to help us with a few interactive scribble lines using which our algorithm is going to identify the foreground and the background, after that it will be the algorithms job to obtain a complete segmentation of the foreground from the background image.

The following figures show how an input image gets scribbling from a user with two different colors (which is also going to be input to our algorithm) and the ideal segmented image output.

Scribbled Input Image                                       Expected Segmented Output Image

## The Graph-Cut Algorithm

The following describes how the segmentation problem is transformed into a graph-cut problem:

1. Let’s first define the Directed Graph G = (V, E) as follows:

1. Each of the pixels in the image is going to be a vertex in the graph. There will be another couple of special terminal vertices: a source vertex (corresponds to the foreground object) and a sink vertex (corresponds to the background object in the image). Hence, |V(G)| = width x height + 2.

2. Next, let’s defines the edges of the graph. As obvious, there is going to be two types of edges: terminal (edges that connect the terminal nodes to the non-terminal nodes) and non-terminal (edges that connect the non-terminal nodes only).

3. There will be a directed edge from the terminal node source to each of non-terminal nodes in the graph. Similarly,  a directed edge will be there from each non-terminal node (pixel) to the other terminal node sink. These are going to be all the terminal edges and hence, |E_T(G)| =  2 x width x height.

4. Each of the non-terminal nodes (pixels) are going to be connected by edges with the nodes corresponding to the neighboring pixels (defined by 4 or 8 neighborhood of a pixel).  Hence, |E_N(G)| =  |Nbd| x width x height.

2. Now let’s describe how to compute the edge weights in this graph.

1. In order to compute the terminal edge weights, we need to estimate the feature distributions first, i.e., starting with the assumption that each of the nodes corresponding to the scribbled pixels have the probability 1.0 (since we want the solution to respect the regional hard constraints marked by the user-seeds / scribbles) to be in foreground or background object in the image (distinguished by the scribble color, e.g.), we have to compute the probability that a node belongs to the foreground (or background) for all the other non-terminal nodes.

2. The simplest way to compute $P_F$ and $P_B$ is to first fit a couple of  Gaussian distributions on the scribbles by computing the parameters (μ, ∑)
with MLE from the scribbled pixel intensities and then computing the (class-conditional) probabilities from the individual pdfs (followed by a normalization) for each of the pixels as shown in the next figures. The following code fragment show how the pdfs are being computed.

```import numpy as np
from collections import defaultdict

def compute_pdfs(imfile, imfile_scrib):
'''
# Compute foreground and background pdfs
# input image and the image with user scribbles
'''
rgb = mpimg.imread(imfile)[:,:,:3]
yuv = rgb2yiq(rgb)
rgb_s = mpimg.imread(imfile_scrib)[:,:,:3]
yuv_s = rgb2yiq(rgb_s)
# find the scribble pixels
scribbles = find_marked_locations(rgb, rgb_s)
imageo = np.zeros(yuv.shape)
# separately store background and foreground scribble pixels in the dictionary comps
comps = defaultdict(lambda:np.array([]).reshape(0,3))
for (i, j) in scribbles:
imageo[i,j,:] = rgbs[i,j,:]
# scribble color as key of comps
comps[tuple(imageo[i,j,:])] = np.vstack([comps[tuple(imageo[i,j,:])], yuv[i,j,:]])
mu, Sigma = {}, {}
# compute MLE parameters for Gaussians
for c in comps:
mu[c] = np.mean(comps[c], axis=0)
Sigma[c] = np.cov(comps[c].T)
return (mu, Sigma)
```

3. In order to compute the non-terminal edge weights, we need to compute the similarities in between a pixel node and the nodes corresponding to its neighborhood pixels, e.g., with the formula shown in the next figures (e.g., how similar neighborhood pixels are in RGB / YIQ space).

3. Now that the underlying graph is defined, the segmentation of the foreground from the background image boils down to computing the min-cut in the graph or equivalently computing the max-flow (the dual problem) from the source to sink.

4. The intuition is that the min-cut solution will keep the pixels with high probabilities to belong to the side of the source (foreground) node and likewise the background pixels on the other side of the cut near the sink (background) node, since it’s going to respect the (relatively) high-weight edges (by not going through the highly-similar pixels).

5. There are several standard algorithms, e.g., Ford-Fulkerson (by finding an augmenting path with O(E max| f |) time complexity) or Edmonds-Karp (by using bfs to find the shortest augmenting path, with O(VE2) time complexity) to solve the max-flow problem, typical implementations of these algorithms run pretty fast, in polynomial time in V, E.  Here we are going to use a different implementation (with pymaxflow) based on Vladimir Kolmogorov, which is shown to run faster on many images empirically in this paper.

## Results

The following figures / animations show the interactive-segmentation results (computed probability densities, subset of the flow-graph & min-cut, final segmented image) on a few images,  some of them taken from the above-mentioned courses / videos, some of them taken from Berkeley Vision dataset.

Input Image

Input Image with Scribbles

Fitted Densities from Color Scribbles

A Tiny Sub-graph with Min-Cut

Input Image

Input Image with Scribbles

Fitted Densities from Color Scribbles

A Tiny Sub-graph with Min-Cut

Input Image (liver)

Input Image with Scribbles

Fitted Densities from Color Scribbles

Input Image

Input Image with Scribbles

Fitted Densities from Color Scribbles

A Tiny Sub-graph of the flow-graph with Min-Cut

Input Image

Input Image with Scribbles

Input Image

Input Image with Scribbles

Fitted Densities from Color Scribbles

A Tiny Sub-graph of the flow-graph with Min-Cut

Input Image

Input Image with Scribbles

A Tiny Sub-graph of the flow-graph with Min-Cut

Input Image                                                         Input Image with Scribbles

Input Image

Input Image with Scribbles

A Tiny Sub-graph of the flow-graph with Min-Cut

Input Image

Input Image with Scribbles

Fitted Densities from Color Scribbles

Input Image

Input Image with Scribbles

Fitted Densities from Color Scribbles

A Tiny Sub-graph of the flow-graph with Min-Cut

Input Image (UMBC Campus Map)

Input Image with Scribbles

Input Image

Input Image with Scribbles

A Tiny Sub-graph of the flow-graph with Min-Cut (with blue foreground nodes)

## Changing the background of an image (obtained using graph-cut segmentation) with another image’s background with cut & paste

The following figures / animation show how the background of a given image can be replaced by a new image using cut & paste (by replacing the corresponding pixels in the new image corresponding to foreground), once the foreground in the original image gets identified after segmentation.

Original Input Image

New Background