Generate a cosine distance heatmap based on CIELAB color space.
How to use:
- Select a picture from the drop down menu;
- Draw a polygon with your mouse: hold left mouse button down to draw, release when finished;
- Wait while it computes the similarity heatmap. Computation is asynchronous, so doesn’t freeze the browser;
- Win win win.
- The histogram is generated based on multiplying the A & B channels from CIELAB.
- CIELAB has two color description channels, compared to HSV’s single color channel. I’m hoping two channels gives a “wider” distribution - while still remaining reasonably invariant to luminosity.
- The distance metric is cosine (i.e., inner product between histograms).
- A sliding window approach is used to generate local histograms describing the surroundings of each pixels. At the time of writing this is 9 by 9 pixels, with the current pixel as center. Pixels are repeated at edges to avoid out-of-bound issues.
- The histogram, at time of writing, has 100 buckets.
The paintings are from WikiArt and the photos from VGG.