FileFormat.md

Input/Output files format

GMM

Input

D k n
α₁
...
α_k
μ_1,1 ... μ_D,1
...
μ_1,k ... μ_D,k
q_1,1 ... q_D,1 l_1,1 ... l_{½D(D - 1),1}
...
q_1,k ... q_D,k l_1,k ... l_{½D(D - 1),k}
x_1,1 ... x_D,1
...
x_1,n ... x_D,n
γ m

Definitions of all variables are given in the srajer-autodiff-screen.pdf, page 3. Note that if replicate point mode is enabled the benchmark expects only a line containing x_1,1 ... x_D,1 and duplicates that point n times.

Output

1. ..._F_... file
   Contains only the value of the function in the specified point.
2. ..._J_... file
   v₁ ... v_n where v_i are components of the objective gradient.

BA

Input

n m p
p₁ ... p₁₁
x₁ x₂ x₃
w₁
feat₁ feat₂

n,m,p are number of cams, points and observations. Definitions of all other variables are given in the srajer-autodiff-screen.pdf, page 4.

Output

1. ..._F_... file

   Reprojection error:
   reproj_err₁
   ...
   reproj_err_2p
   Zach weight error:
   w_err₁
   ...
   w_err_p
   

2. ..._J_... file

   This file contains sparse Jacobian of nrows * ncols size in the CSR format where ncols=2p+p, nrows=11n+3*m+p. It suggests the use of three one-dimensional arrays (rows,cols,vals).

   • vals holds all the nonzero entries of the Jacobian in the left-to-right top-to-bottom order.
   • rows is of length nrows + 1. It is defined recursively as follows:
     rows[0] = 0
     rows[i] = rows[i-1] + the number of nonzero elements on the i-1 string of the Jacobian
   • cols[i] contains the column index in the Jacobian of each element of vals and that's why it has the same size

   The resulting file looks as follows:
   nrows ncols
   rows_size
   rows₀ ... rows_{rows_size-1}
   cols_size
   cols₀ ... cols_{cols_size-1}
   vals₀ ... vals_{rows_size-1}
   

Hand

Input

1. model/bones.txt Contains a list of lines where each line contains such parameters separated by ":" delimiter:

   • bone_name
   • bone_parent
   • base_relative₁ ... base_relative₁₆
   • base_absolute₁ ... base_absolute₁₆

2. model/vertices.txt Contains a list of lines where each line contains such parameters separated by ":" delimiter:

   • v₁ ... v₃
   • dummy₁ ... dummy₅
   • n
   • bone₁:weight_bone1,vert1: ... :bone_n:weight_bonen,vertn

3. model/triangles.txt Contains a list of lines:

   • v₁:v₂:v₃

4. input.txt

   N n_theta
   correspondence₁ point_1,1 point_1,2 point_1,3
   ...
   correspondence_N point_N,1 point_N,2 point_N,3
   u_1,1 u_1,2
   ...
   u_{n_pts,1} u_N,2
   θ₁
   ...
   θ_{n_theta}
   

Note that the benchmark expects "u" block only if complicated mode is enabled.

Definitions of all variables are given in the srajer-autodiff-screen.pdf, page 5.

Output

1. ..._F_... file
   v₁ ... v_n where v_i are components of the objective vector.

2. ..._J_... file

   j_1,1 ... j_1,ncols
   ...
   j_nrows,1 ... j_nrows,ncols
   

   where ncols=(complicated ? 2 : 0) + n_theta, nrows=3*n_pts

LSTM

Input

l c b
main_param₁ ... main_param_2l·4b
extra_param₁ ... extra_param_3b
state₁ ... state_2l·b
seq₁ ... seq_c·b

Output

1. ..._F_... file
   Contains only the value of the function in the specified point.
2. ..._J_... file
   v₁ ... v_n where v_i are components of the objective gradient.