jvlmdr

---------------------------------------------------------------------------------- benchmark 'test_benchmark_assign_3x3': 6 tests ---------------------------------------------------------------------------------
Name (time in us)                             Min                 Max                Mean             StdDev              Median                IQR            Outliers  OPS (Kops/s)            Rounds  Iterations
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_benchmark_assign_3x3[lapsolver]      21.8828 (1.0)      382.6180 (1.0)       27.1516 (1.0)      13.0945 (1.0)       23.1010 (1.0)       1.3541 (1.0)     1014;2011       36.8302 (1.0)       11318           1
test_benchmark_assign_3x3[lap]            41.3780 (1.89)     887.3888 (2.32)      52.0903 (1.92)     23.1687 (1.77)      45.7240 (1.98)      1.9629 (1.45)     853;1701       19

jvlmdr

#!/bin/bash

baseurl="http://cvlab.hanyang.ac.kr/tracker_benchmark"
wget "$baseurl/datasets.html"
cat datasets.html | grep '\.zip' | sed -e 's/\.zip".*/.zip/' | sed -e s'/.*"//' >files.txt
cat files.txt | xargs -n 1 -P 8 -I {} wget -c "$baseurl/{}"

jvlmdr

import numpy as np

def covar1(m, s):
    s = float(s)
    r = np.arange(m, dtype=np.float)
    d = np.reshape(r, (m, 1)) - np.reshape(r, (1, m))
    k = np.exp(-(0.5/s**2) * (d * d))
    return k

def smooth1(m, s):

jvlmdr

function benchmark()
    test();
    m1 = 32;
    m2 = 48;
    k = 24;
    x = randn(m1, m2, k);
    af = covar1(randn(m1, m2, k));
    a = ifft2(af, 'symmetric');

    t = time_func(@() mul1(af, x));

jvlmdr

% Computes circulant covariance of a and then multiplies x by covariance.
% Both images are of size (m1, m2) with k channels.
%
% size(a) is [m1, m2, k]
% size(x) is [m1, m2, k]
function b = mul_circ_covar(a, x)
    [m1, m2, k] = size(x);

    % compute circulant covariance from shifts of a
    % sf(u1,u2,p,q) = conj(af(u1,u2,q)) * af(u1,u2,p)

jvlmdr

function [x, mul_dx_dA, mul_dx_db] = solve_rect(A, b)
% solve_rect returns x that minimises |A x - b|^2
%   x = (A' A)^-1 A' b
% and operators that compute derivatives with respect to A and b.
% It uses a QR decomposition of A.
%
% Parameters:
% A has size [m, n] with m >= n and rank(A) = n.
% b has size [m, 1].
%

jvlmdr

function [x, mul_dx_dA, mul_dx_db] = solve_square(A, b)
% solve_square returns x = A^-1 b and operators that compute products with
% derivatives with respect to A and b. It uses an LU decomposition of A.
%
% Parameters:
% A has size [n, n] and rank(A) = n.
% b has size [n, 1].
%
% Returns:
% x has size [n, 1].

jvlmdr

function dy_dx = deriv_lsq_nonlin(rfun, x, y_hat, m, n)
% deriv_lsq_nonlin estimates the derivative of
%   y(x) = argmin_y |r(x, y)|^2 .
%
% Parameters:
% [r, dr_dx, dr_dy, d2r_dxdy] = rfun(x, y)
%   x has size [m, 1].
%   y has size [n, 1].
%   r has size [p, 1].
%   dr_dx has size [p, m] and dr_dx(i, j) = dr(i) / dx(j).

jvlmdr

n = 100;
m = 50;
A = 1/sqrt(m) * randn(m, n);

K = 10;
rmin = ones(K, 1);
rmax = ones(K, 1);
trials = 1e4;
for k = 1:K
  for i = 1:trials

jvlmdr

n = 100;
m = 50;
A = 1/sqrt(m) * randn(m, n);

K = 10;
trials = 1e4;
rmin = ones(K, 1);
rmax = ones(K, 1);
for k = 1:K
  for i = 1:trials