Complete Classification Results

For a proper test of the trained neural network, we splitted our samples by users (instead of using the 0.7 / 0.3 approach).

Training data:
User 0 = 498 samples
User 1 = 510 samples
User 2 = 504 samples
User 3 = 524 samples

Total of 2035 samples (1425 for training + 610 for validation)
 
Test data:
User 4 = 261 samples
User 5 = 261 samples
 
Total of 522 samples 

This way, we can ensure that all test samples were collected at a different time (with new force/torque sensor calibration) and by different users than the training samples.


Learning Curves


 
Epoch 264/300
loss: 0.3699 - accuracy: 0.9846
val_loss: 0.4017 - val_accuracy: 0.9705
 
Using 1425 samples for training and 610 for validation

 

Confusion Matrices 

Total test samples: 522
Correctly predicted: 507
 

 

 


 

 

 

 

Outputs Confidence Analysis 


 

 

 

 

Legend: 
  • Max confidence reached (green line)
  • Mean of every confidences (bold value)
  • Standard deviation (right below mean)
  • Min confidence reached (red line

 

Accuracy Analysis per test User 

User 4

Confusion Matrices

 

Total test samples: 261
Correctly predicted: 248


 

 


Outputs Confidence Analysis

  

 

 

 

Legend: 
  • Max confidence reached (green line)
  • Mean of every confidences (bold value)
  • Standard deviation (right below mean)
  • Min confidence reached (red line

 

User 5

Confusion Matrices

Total test samples: 261
Correctly predicted: 259

 

 

 

Outputs Confidence Analysis

 

 

 

 

 

Legend: 
  • Max confidence reached (green line)
  • Mean of every confidences (bold value)
  • Standard deviation (right below mean)
  • Min confidence reached (red line)

 

Comments

Popular posts from this blog

Real-time UR10e following a tracked object

RGB-D tracking + UR10e following & picking/placing

Remote Control of UR10e via MoveIt (ROS)