Skip to content

Latest commit

 

History

History

README.md

FaceRec Evaluation

Ground Truth creation

We realised two ground truth for evaluating the system.

Approach:

  • Create a list of people and train the system on them
  • Retrieve from the Knowledge Graphs the shots/segments in which these people should appear.
  • Run a first face recognition, in order to artificially realise a subset of appearences of each person.
  • We manually correct the obtained dataset, adding also some shots/segments not involving any of the specified people.

Notebooks for the generation:

Some differences between the two datasets:

ANTRACT MeMAD
type historical images TV news
years 1945-1969 2014
resolution 512x384 455x256
colorspace b/w colour
shots division yes no
list of celebrities to search 13 (chosen by domain experts) 6 (most present in KG)
represented fragment and length shot, 3 seconds in avg. segment, up to 2 minutes
records 216 100
distinct fragments 198 100
distinct media 129 30
nb. fragments without known faces 39 43

Evaluation

We run face recognition on the two datasets and measure the matches with the expected people.

Notebooks for the evaluation:

Some results:

ANTRACT Precision Recall F-Score Support
Ahmed Ben Bella 1.00 0.46 0.63 13
François Mitterrand 1.00 0.92 0.96 13
Pierre Mendès France 1.00 0.61 0.76 13
Guy Mollet 0.92 0.92 0.92 13
Georges Bidault 0.83 0.71 0.76 14
Charles De Gaulle 1.00 0.57 0.73 19
Nikita Khrouchtchev 1.00 0.38 0.55 13
Vincent Auriol 1.00 0.46 0.63 13
Konrad Adenauer 1.00 0.53 0.70 13
Dwight Eisenhower 0.85 0.46 0.60 13
Elisabeth II 1.00 0.71 0.83 14
Viatcheslav Molotov 1.00 0.23 0.37 13
Georges Pompidou 1.00 0.69 0.81 13
-- unknown -- 0.35 0.97 0.52 39
average (unknown apart) 0.97 0.59 0.71 216
MeMAD Precision Recall F-Score Support
Le Saint, Sophie 0.90 0.90 0.90 10
Delahousse, Laurent 1.00 1.00 1.00 7
Lucet, Elise 1.00 0.90 0.94 10
Gastrin, Sophie 1.00 0.90 0.94 10
Rincquesen, Nathanaël de 1.00 0.80 0.88 10
Drucker, Marie 1.00 1.00 1.00 10
-- unknown --- 0.89 0.97 0.93 43
average (unknown apart) 0.98 0.91 0.94 100