Computer Vision News - December 2016

Essentially, the reducibility condition calls for smooth changes in the values of the neighboring pixels. From this perspective, this is an intuitive requirement on the minimum sampling density that is needed for tracking local changes in the shape boundaries. Results: Computer Vision News Research 33 Research To illustrate some of the algorithm’s capabilities, the authors took shape images at the resolution 1000x1000, then created 200x200 discrete images of them with biquadratic B-spline sampling kernels; successively, they reconstructed continuous shapes and compared them to the originals. The figure above demonstrates this process: (a) and (c) are the original images; (b) and (d) are an enlarged section of each, on which the rest of the illustration focuses; (e) and (g) are the respective discrete images; (f) and (h) are the reconstructed shapes. In the figure on the right, (i) is a 200x200 discrete image corresponding to (a) above, which was created with sampling kernels that are shifts of a stretched biquadratic B-spline with an effective support of 40x40 pixels; (j) is the corresponding enlarged section, as above, and (k) is the reconstructed image. The recovered image (without any thresholding) has PSNR of 33.8096 dB with respect to the original shape and a measurement PSNR of 75.0489 dB. The multi-constraint minimization problem is equivalent to the Cheeger problem