Multiple Feature Fusion Based on Co-Training Approach and Time Regularization for Place Classification in Wearable Video

Algorithm 1

The CO-DAS and CO-TA-DAS algorithms.

INPUT:

 Training set 𝐿=𝐱_𝑖,𝑦_𝑖^𝑙_𝑖=1;

 Testing set 𝑈=𝐱_𝑖^𝑢_𝑖=1;

OUTPUT:

 𝑦_𝑖—class estimations for the testing set 𝑈;

 𝑓⁽¹⁾, 𝑓⁽²⁾—trained classifiers;

PROCEDURE:

 (1) Compute visual features 𝐱_𝑖=(𝐱_𝑖⁽¹⁾,𝐱_𝑖⁽²⁾) for every image 𝐼_𝑖 in the dataset

 (2) Initialize 𝐿₁=𝐱_𝑖⁽¹⁾,𝑦_𝑖^𝑙_𝑖=1 and 𝐿₂=𝐱_𝑖⁽²⁾,𝑦_𝑖^𝑙_𝑖=1

 (3) Initialize 𝑈₁=𝐱_𝑖^(1)𝑢_𝑖=1 and 𝑈₂=𝐱_𝑖^(2)𝑢_𝑖=1

 (4) Create two work sets 𝑈₁∶=𝑈₁ and 𝑈₂∶=𝑈₂

 (5) Repeat until the sets 𝑈₁ and 𝑈₂ are empty (CO)

       (a) Train classifiers 𝑓⁽¹⁾, 𝑓⁽²⁾ using the sets 𝐿₁, 𝐿₂ respectively;

       (b) Classify the patterns in the sets 𝑈₁ and 𝑈₂ using the classifiers 𝑓⁽¹⁾ and 𝑓⁽²⁾ respectively;

     (i) Compute scores 𝑠⁽¹⁾_test and confidences 𝑧⁽¹⁾ on the set 𝑈₁

     (ii) Compute scores 𝑠⁽²⁾_test and confidences 𝑧⁽²⁾ on the set 𝑈₂

       (c) Add the 𝑘 top confidence estimations 𝐿₁⊂𝑈₁, 𝐿₂⊂𝑈₂

           (i) 𝐿₁∶=𝐿₁∪𝐿₁

           (ii) 𝐿₁∶=𝐿₁∪𝐿₁

       (d) Remove the 𝑘 top confidence patterns from the working sets

     (i) 𝑈₁∶=𝑈₁⧵𝐿₁

     (ii) 𝑈₂∶=𝑈₂⧵𝐿₂

       (e) Go to step (5).

 (6) Optionally: perform Temporal Accumulation (TA) according to (15)

 (7) Perform classifier output fusion (DAS)

       (a) Compute fused scores 𝐬^DAS_test=1−𝛽𝐬⁽¹⁾_test+𝛽𝐬⁽²⁾_test;

        (b) Output class estimations 𝑦_𝑖 from the fused scores 𝐬^DAS_test