Skip to content

Latest commit

 

History

History
41 lines (32 loc) · 2.61 KB

README.md

File metadata and controls

41 lines (32 loc) · 2.61 KB
Raw

What is our project?

We aim to reunite children with their parents after disaster. By equipping every shelter / disaster relief site with a node in the network and a camera, we use kinship verification technology to match children with their mother and father.

Hypothesis

The hypothesis of this project is that the age-invariant features of children are inherited from parents. By inferring from these features, we may tell whether a child is related to a father-mother pair.

Architecture

We utilize transfer learning by extracting age-invariant features from the mother, father and child. We learn a comparison function by the use of fully-connected and concatenation layer. Our approach is to compare the father and child, and the mother and child. Then, these comparisons are concatenated, and are inferred from to obtain a final relatedness score (0-1).

Network Architecture

results

Date Dataset Model Validation Acc Validation Loss Training Acc starting LR Epochs Batch size Batches per epoch Notes Weights filename
Sep 5 2019 FIW model_v3 90.26% 0.389 91.37% 0.0001 2410 50 F-M-C(+)-C(-) pairs 20 Reduced LR on plateau after a patience of 75 epochs src/ml/saved_model_weights/v3_weights.2409-0.39.hdf5
Sep 1 2019 FIW model_v3 86.57% 0.495 99.05% 0.0001 2134 50 F-M-C(+)-C(-) pairs 20 Reduced LR on plateau after a patience of 75 epochs src/ml/saved_model_weights/v3_weights.2134-0.50.hdf5

Note: to use the model weights, you must point the model saver (see train_model_v3.py) to the location of the weights.

TODO

Things we need to do:

  • Implement more loss functions if needed (e.g. crossentropy?)
  • Visualize using tensorboard.
  • Finish train_model.py and model.py
  • Implement data loader class to load images from FIW / other dataset
  • Implement algorithm to sample (MF-C) pairs used in triplet pairs
  • Fix training bug in tensorflow - (multiple forward passes result in weight update)
  • Try reducing LR on plateau after 80 epochs

Papers used:

Look Across Elapse

 @article{zhao2018look,
      title={Look Across Elapse: Disentangled Representation Learning and Photorealistic Cross-Age Face Synthesis for Age-Invariant Face Recognition},
      author={Zhao, Jian and Cheng, Yu and Cheng, Yi and Yang, Yang and Lan, Haochong and Zhao, Fang and Xiong, Lin and Xu, Yan and Li, Jianshu and Pranata, Sugiri and others},
      journal={AAAI},
      year={2019}
      }