uf_tree indexing fix

[greenlet]

No description provided.

[greenlet]

I measured performance like in the code for the sceaux_global_reconstruction set:

1. develop baseline gives 3.5 sec
2. issue-1737-experimental with std::map<indexedFeaturePair, uint32_t> image_feature_index instead of std::set<indexedFeaturePair> allFeatures plus flat_pair_map<indexedFeaturePair, uint32_t> map_node_to_index gives 4.2 sec. So std::map didn't work out
3. Current issue-1737 with std::unordered_set<indexedFeaturePair, image_feature_hash> allFeatures and flat_pair_map<indexedFeaturePair, uint32_t gives 2.8 sec. Also uf_tree indexing fixed
   https://github.com/openMVG/openMVG/pull/1740/files#diff-cb5e7e1f5559c0a219a2f49376872a98R151
   https://github.com/openMVG/openMVG/pull/1740/files#diff-cb5e7e1f5559c0a219a2f49376872a98R199

[greenlet]

Is it possible to have two similar indexedFeaturePair at all?

[pmoulon]

We should not since we are using a set

[pmoulon]

I made some benchmark on my machine and I found actually that your version is a bit slower.

Here the results on various datasets.

Dataset	Time Old(s)	Time New(s)	%
chambord	12.3622	15.4362	.8
Dataset_Saint_Lys/Eglise	5.17	5.88403	.87
2017/LaPush/RubyBeach	0.518462	0.658825	.78
2016/Eglise_Cognat	17.0268	20.2247	.85

So on my machine, your code is in average slower (20%)

[greenlet]

I made some benchmark on my machine and I found actually that your version is a bit slower.

Here the results on various datasets.

Dataset Time Old(s) Time New(s) %
chambord 12.3622 15.4362 .8
Dataset_Saint_Lys/Eglise 5.17 5.88403 .87
2017/LaPush/RubyBeach 0.518462 0.658825 .78
2016/Eglise_Cognat 17.0268 20.2247 .85
So on my machine, your code is in average slower (20%)

What targets your test included? Please note that in current version there is a cycle which should be removed before testing main_exportTracks:
https://github.com/openMVG/openMVG/pull/1740/files#diff-278208fdb61f1a04029842126484e631R129-R137

[pmoulon]

It could due to the compiler, I will test on another machine to see the impact.
Good point: Memory consumption was identical for both algorithms