IdentityFinder | Devpost

Inspiration

Each year, there are over 30,000 reports of missing children sent to police in Canada. Unfortunately, not all of them are able to return to their own families. When these children grow up, they may want to find their parents or want to know more about themselves. However, it is not an easy task to do as there are too many cases and hence nearly impossible to do it manually. Therefore, we want to provide a solution through this project, to help them find their homes.

What it does

The application will identify whether a photo of a grown-up and a photo of a child illustrate the same person.

How we built it

We transformed image data into tensors, then we applied CNN(Convolutional Neural Network) to train the model.

Challenges we ran into

It is not that easy to find people's photos at their different ages due to the strict Child Protection Law. We finally found a paper analyzing photos of famous stars at their different ages: link

It also was a challenge to compare between 2 images. The common task of computer vision project to build classifiers for images, for which the label/target will be a number/category or a list of labels. In our case, the target is an image and we spent a long time choosing the appropriate loss function and deciding what structure would the input and target be.

Since image data are large-scaled, it took much longer to train the model than using simple numerical data. Not to mention we only had our laptops to run the training.

Accomplishments that we're proud of

We finally achieve an average loss of around 2% for training, validation and test sets.

What we learned

We had a deeper understanding of applying CNN, especially training size deciding, layer info calculation and hyper-parameters tuning.

What's next for IdentityFinder

Applying different deeper networks is a must to do further improvements in terms of identification accuracy; Finding more photos as test cases is required as well. We will try to collect photos using other datasets or even from our friends and families. After the model is well adjusted, we will post it onto our website for public use, or provide it to charity organizations.