Content-aware image scaling as described by [1].
Usage example: python carve.py broadway_tower.jpg output/broadway_tower --iteration_count=300
To see the carving animated, use the make_gif.py script.
Usage example: python make_gif.py output/broadway_tower
Original image:
Energy function (e.g. gradient magnitude):
Lowest energy seam (e.g. using Dijkstra's algo):
Shrunk image (297 iterations):
width=3440px
width=2440px
width=3440px
width=2440px
width=1440px
Supported energy functions:
- gradient magnitude
- spectral residual saliency
- fine grained saliency
- spectral saliency + gradient magnitude
- entropy (3-channel or grayscale)
Entropy tends to be the slowest, spectral saliency is the fastest. Gradient usually perform the best. Due to the overhead in conversion, 3-channel entropy is faster than grayscale.
[1] Avidan, Shai; Shamir, Ariel (July 2007). "Seam carving for content-aware image resizing | ACM SIGGRAPH 2007 papers". Siggraph 2007: 10. doi:10.1145/1275808.1276390