d31 entity recognition results with auto- associative memories nicolas gourier inria prima team...
TRANSCRIPT
D31 Entity RecognitionResults with Auto-
associative Memories
Nicolas GourierINRIAPRIMA TeamGRAVIR
Laboratory
CAVIAR Project
Entity recognition
Can be performed either by local or global approaches
Local approaches Use information contained in the
neighboorhood of pixels
Global approaches Use the entire appearance of the
image
Why a global approach ?
No landmarks have to be detected
No model has to be constructed
Can handle Low resolution Partial occlusions
=> Only the object has to be detected
Existing Global Approaches
PCA, KDA,… [Pentland91] Sensitive to alignment Number of dimensions ?
Neural networks Number of cells in the hidden layer ? Recovery of prototypes of image
classes ?
=> Auto-associative Memories [Abdi94]
Plan of the talk
1) Our approach
2) Results
3) Comparison with other techniques
1. Normalized object imagette Grey scale face imagette normalized in
size and slant: 25x25 pixels
=> Computation time reduction=> Size and slant robustness
1. Auto-associative memories
Linear auto-associative memory
Input patterns associated with themselves
Connection between input unitsPortion of an input =>Complete pattern
X’ = W.X X : Source image X’ : Output image W : Weights
1. Hebbian learning rule
W = Xk.XkT
Faces not well discriminated [Valentin94]
1. Widrow-Hoff learning rule (1)
Learned images are reconstructed
Other images are degraded [Valentin94]
1. Widrow-Hoff learning rule (2)
Creation of prototypes
Eigenvalues egalization [Abdi & Valentin94]
We adapt Widrow-Hoff learning to entity recognition
1. Entity Recognition
Compare the input image to all responses=> Score between 0 and 1
Winner-takes-all process
-> ½ videos for training,½ videos for test
1. Training and Test
Training ->Test V
2. Experiments
3 Experiments :
1) Classes 0 / 1 person Without training a 0 person class
2) Classes 0 / 1 person
3) Classes 0 / 1+ person
2. Result of the first experiment (1)
2. Result of the first experiment (2) Not sufficient for reliable
classification
0 person class imagettes have non-uniform variations in appearance
=> Learn a 0 person class from random images of the background
2. Result of the second experiment
2. Result of the third experiment
2. Recall and precision
Experiment
Classes
11/0
21/0
31+/0
1st class recall
- 99 % 99 %
2nd class recall
- 68 % 70 %
1st class precision
- 95 % 93 %
2nd class precision
- 93 % 90 %
3. Discussion
Training the 0 person class improves discrimination
Some 0 person class images are misclassified :
3. Advantages Varying the size of the imagette do not
have much influence -> 25x25 pixels
Normalization + Classification is done at video-rate
Prototypes can be saved and reused
Can be adapted to entity recognition
Conclusion
+ Invariance to scale, slant and alignment
+ Not disrupted by local changes
- Needs to train a non-person class
Adapted to the project Low resolution Changes of viewpoint Fast