The goal of this project was to recreate an image as a mosaic of other images. I wanted to achieve this with a low number of tiles and not simply use images as pixels based on their colors.
The biggest mistake I made was to vastly underestimate the difficulty of the problem. I thought I could get good results by tuning my early histogram grid implementation and floundered. I have seen many other attempts at this project but yet to see any that produce good results.
Path to Failure
- Played with Perceptual Hashes (then read about how they work and abandoned this path).
- Moved to Grid of Histograms: Divide source and input images into grids and generate / compare Histograms.
- Switched from HSV to LAB / LUV color spaces for a more human perception of color.
- Switched from Instagram to Flickr for source images(Insta is filled with images of text).
- Experimented with different histogram comparison methods (Chi-Square, Bhattacharyya, etc).
- Optimised algorithm for a shorter iteration loop (this was successful).
- Researched our perception of shapes, switched to basic shape (still factor in color) approach.
- Performance mandated hybrid approch: first pass histogram grid, second pass with feature detection, finally adjust image contrast / brightness.
- Quantify success early: manually create mosaic based on small input set and diff with algorithms results to calculate how successful each run was.
- Search inside each possible image for good matches / try different rotations of source images.
- More processing power (cluster).
- Focus on shapes and start in greyscale space.
- Use bag of words model.
- Use weighted multiple factors (histograms, FLANN, Feature Detection, shapes, etc).
- Write it in C++: Other language bindings are great but the stack traces force you into C/C++ anyway.