Please enable JavaScript.
Coggle requires JavaScript to display documents.
High Performance Large Scale Face Recognition with Multi Cognition Softmax…
High Performance Large Scale Face Recognition with Multi Cognition Softmax and Feature Retrieval
MCSM Module
Using balanced and cleaned large scale data
Training data is distributed to several cognition units by a data shuffling
Results from each cognition unit are merged with a :arrow_right:
score-level average
max-max strategy
Targeted training data of each cognition unit is divided into several small subsets with independent softmax model
Other Softmax Models
Softmax Model
Targeted mapping will be more complex
As complexity increases,
needed deeper convolutional layers,
therefore heavier computational resources
Misclassification risk may be increased
Independent Softmax Model
Predict the probabilities for a part of classes
Final results are computed by a max-max scheme
Conflict happens between ISMs
MCSM combines the power of SM and ISM (divide and conquer)
Feature Retrieval Module
Template based method
Average the features extracted by DCNN
Each targeted identity is fully represented in a low dimensional feature
Voting scheme
Primary model with best performance
Auxiliary models
If all of models are the same but different from primary:
we replace the predicted label and score with these models' prediction + averaging score plus an offset p
One-shot Learning Module
Only testing images with low confidence score from MCSM & FRetrieval
Assigned with new labels with higher score by merging one-shot learning result
Cleaned Data
Step 1
Semantic bootstrapping.
If one image of a celeb is > threshold then defined as "Clean image"
until satisfies the number of image as "Clean Celebrity"
Step 2
Feature-based clustering method to clean non-"Clean Celebrity"
using
cosine similarity
with Top-5 ranked image with thres. 0.5
Step 3
Manually add more images of the remaining celebrity to balance the data