This repository contains an implementation of the fully convolutional Tiramisu network for semantic segmentation, applied to the task of playing Where's Waldo?.
Where’s Waldo (or Wally) is a series of children’s books that challenges the reader with finding the eponymous character and his compadres in densely illustrated images.
Please see my recent blog post Where's Waldo: Terminator Edition for an overview of my approach.
I've also provided a complete end-to-end walk-through of my There's Waldo project in this notebook.
You can find all training / testing images and corresponding predictions here
I've provided for you the original images and bounding box xml's necessary to recreate my process.
First, unpack them:
$ unzip imgs_boxes.zipTo create the binary labels:
$ python make_targets.pyThe next step is to create the Waldo sub-images.
$ python create_subimages.pyI didn't put together a script for training the model; training neural networks isn't straightforward and relies alot on human judgement. If you'd like to train your own model, please take a look at the accompanying notebook.
To make predictions on new images:
$ python predict.py image1.py image1.py --output_path /output/pathThis will load weights from my model by default. Makes predictions and saves them as transparency masks overlaid on the original image. They'll be saved as output_{i}.png in the specified folder.
If you want to use your own model weights, just specify with the --model flag.
This project is inspired by material studied while interning with Jeremy Howard and Rachel Thomas' fast.ai, in particular Lesson 14 of their course Cutting Edge Deep Learning for Coders, taught at USF’s Data Institute.