by Antonin Kenens and Tangi Migot
A Julia repository of bundle adjustment problems from the Bundle Adjustment in the Large repository.
using BundleAdjustmentModels, DataFrames
The function problems_df() returns a Dataframe of all the bundle adjustment problems.
There are 74 different models organized in the Dataframe with the following elements:
the name of the model;
the group of the model;
the size of the jacobian matrix (nequ, nvar);
the number of non-zeros elements in the jacobian matrix (nnzj).
df = problems_df()
74×5 DataFrame
Row │ name group nequ nvar nnzj
│ String String Int64 Int64 Int64
─────┼──────────────────────────────────────────────────────────────────
1 │ problem-16-22106-pre dubrovnik 167436 66462 2009232
2 │ problem-88-64298-pre dubrovnik 767874 193686 9214488
3 │ problem-135-90642-pre dubrovnik 1106672 273141 13280064
4 │ problem-142-93602-pre dubrovnik 1131216 282084 13574592
5 │ problem-150-95821-pre dubrovnik 1136238 288813 13634856
6 │ problem-161-103832-pre dubrovnik 1184038 312945 14208456
7 │ problem-173-111908-pre dubrovnik 1269140 337281 15229680
8 │ problem-182-116770-pre dubrovnik 1337410 351948 16048920
⋮ │ ⋮ ⋮ ⋮ ⋮ ⋮
68 │ problem-1158-802917-pre venice 8261006 2419173 99132072
69 │ problem-1184-816583-pre venice 8358094 2460405 100297128
70 │ problem-1238-843534-pre venice 8581006 2541744 102972072
71 │ problem-1288-866452-pre venice 8766012 2610948 105192144
72 │ problem-1350-894716-pre venice 9034252 2696298 108411024
73 │ problem-1408-912229-pre venice 9269060 2749359 111228720
74 │ problem-1778-993923-pre venice 10003892 2997771 120046704
59 rows omitted
For instance, it is possible to select the problems where the Jacobian matrix of the residual has at least 50000 lines and less than 34000 columns.
filter_df = df[ ( df.nequ .≥ 50000 ) .& ( df.nvar .≤ 34000 ), :]
2×5 DataFrame
Row │ name group nequ nvar nnzj
│ String String Int64 Int64 Int64
─────┼──────────────────────────────────────────────────────
1 │ problem-49-7776-pre ladybug 63686 23769 764232
2 │ problem-73-11032-pre ladybug 92244 33753 1106928
The Dataframe is listing the matrices that you can have access to, but they still need to be downloaded.
Following the example above, we filtered two problems. What we want to do now is to select the first one in the listing.
name = filter_df[1, :name] # select the name of the first problem
"problem-49-7776-pre"
Now that the name is selected, we need to access the problem itself, and there are 2 solutions:
This package uses Julia Artifacts to handle the problems archives so that
The models are downloaded only once;
They are identified with a unique hash;
They can be deleted with a single command line.
The method fetch_ba_name will automatically download the problem (if needed) and return its path.
path = fetch_ba_name(name)
"/home/runner/.julia/artifacts/dd2da5f94014b5f9086a2b38a87f8c1bc171b9c2"
It is also possible to directly download and get access to an entire group of problems using fetch_ba_group.
paths = fetch_ba_group("ladybug")
30-element Vector{String}:
"/home/runner/.julia/artifacts/dd2da5f94014b5f9086a2b38a87f8c1bc171b9c2"
"/home/runner/.julia/artifacts/3d0853a3ca8e585814697fea9cd4d6956692e103"
"/home/runner/.julia/artifacts/5c5c938c998d6c083f549bc584cfeb07bd296d89"
"/home/runner/.julia/artifacts/e8ebdcec7cd5dc3ec4825dcb6642064e541fa625"
"/home/runner/.julia/artifacts/111ebd1dcdf5670eac8a12f4e2ce5a3b79b4f7f6"
"/home/runner/.julia/artifacts/3e8acd1ef4cb13d93a7140be1c3702b5eb165978"
"/home/runner/.julia/artifacts/69595726ce3990604ef56b80c273fe7f95506f2e"
"/home/runner/.julia/artifacts/1eab751f4850b1e22302a7f97894d4b769c7350e"
"/home/runner/.julia/artifacts/90ce1588d82c76a81c2b07d9e6abcc984e0e4586"
"/home/runner/.julia/artifacts/05e4248f24d897a97e9be1f386e18d1de7e511e5"
⋮
"/home/runner/.julia/artifacts/72cdf832cd3eedb4cc4c589127e7ab5b91a60f00"
"/home/runner/.julia/artifacts/1e01d6927d4d4e0acd3086d9c75d2ee783ab1220"
"/home/runner/.julia/artifacts/b92a8dfd520e7440769c4a72685e817632f9ea43"
"/home/runner/.julia/artifacts/366a8368c091027a23a190b1ed498158eef947a5"
"/home/runner/.julia/artifacts/ee08f4f6898cc26471d3a63502fe908cfa9a5dcc"
"/home/runner/.julia/artifacts/1a0ebf23b35d784ad3d39f4987bc7d785f5976b6"
"/home/runner/.julia/artifacts/00be55410c27068ec73261e122a39258100a1a11"
"/home/runner/.julia/artifacts/0303e7ae8256c494c9da052d977277f21265899b"
"/home/runner/.julia/artifacts/389ecea5c2f2e2b637a2b4439af0bd4ca98e6d84"
Now, it is possible to load the model using BundleAdjustmentModel
df = problems_df()
filter_df = df[ ( df.nequ .≥ 50000 ) .& ( df.nvar .≤ 34000 ), :]
name = filter_df[1, :name]
model = BundleAdjustmentModel(name);
or
model = BundleAdjustmentModel("problem-49-7776-pre");
The function BundleAdjustmentModel will instantiate the model and automatically download it if needed. The resulting structure is an instance of AbstractNLPModel. So, it is possible to access its API as any other NLPModel.
using NLPModels
Using residual, it is possible to compute the residual of the model
model = BundleAdjustmentModel("problem-49-7776-pre.txt.bz2")
x = get_x0(model) # or `model.meta.x0`
Fx = residual(model, x)
63686-element Vector{Float64}:
-9.020226301243213
11.263958304987227
-1.8332297149469525
5.304698960898122
-4.332321480806684
7.117305031392988
-0.5632751791501462
-1.062178017695885
-3.9692059546843836
-2.2850712830954194
⋮
-0.4377574792022081
-0.024461522651677114
-0.1961281328666189
0.008375315306523134
0.23044496991071384
0.04927878647407624
0.47289578243411867
-0.01443314653496941
-0.4486499211288866
or use the in-place method residual!
model = BundleAdjustmentModel("problem-49-7776-pre.txt.bz2")
x = get_x0(model) # or `model.meta.x0`
nequ = get_nequ(model) # or `model.nls_meta.nequ`
Fx = zeros(nequ)
residual!(model, x, Fx);
You can also have access to the LinearOperator of the Jacobian matrix of the residual of the model which is calculated by hand (in contradiction to automatic differentiation).
You need to call jac_structure_residual! at least once before calling jac_op_residual!.
model = BundleAdjustmentModel("problem-49-7776")
meta_nls = nls_meta(model)
nnzj = meta_nls.nnzj # number of nonzeros in the jacobian
rows = Vector{Int}(undef, nnzj)
cols = Vector{Int}(undef, nnzj)
jac_structure_residual!(model, rows, cols);
You need to call jac_coord_residual! to update it to the current point.
model = BundleAdjustmentModel("problem-49-7776")
x = get_x0(model)
vals = similar(x, nnzj)
jac_coord_residual!(model, x, vals)
764232-element Vector{Float64}:
545.1179297695714
-5.058282392703829
-478.0666614182794
-283.5120110272222
-1296.3388697208218
-320.60334752077165
551.1773498438256
0.00020469082949125088
-471.09490058346296
-409.36200783910084
⋮
1246.0491342213195
-96.64898276178089
622.4628407807937
2.852854717671426e-7
305.00859581263086
19.561762292226746
6.5984133899730155
1.680958440896614
0.06413511779846102
Finally you can use jac_op_residual!:
model = BundleAdjustmentModel("problem-49-7776")
meta_nls = nls_meta(model)
nnzj = meta_nls.nnzj
rows = Vector{Int}(undef, nnzj)
cols = Vector{Int}(undef, nnzj)
jac_structure_residual!(model, rows, cols)
vals = similar(model.meta.x0, nnzj)
jac_coord_residual!(model, model.meta.x0, vals)
Jv = similar(model.meta.x0, meta_nls.nequ)
Jtv = similar(model.meta.x0, meta_nls.nvar)
Jx = jac_op_residual!(model, rows, cols, vals, Jv, Jtv)
Linear operator
nrow: 63686
ncol: 23769
eltype: Float64
symmetric: false
hermitian: false
nprod: 0
ntprod: 0
nctprod: 0
There is no second-order information available for the problems in this package.
Once you have finished working with a specific model you can delete it the same way you downloaded it.
delete_ba_artifact!("problem-49-7776")
If you want to clean your workspace, you can also delete all the problems at once.
delete_all_ba_artifacts!()