Retraction
An operation that maps the Gaussian jet onto a hierarchical weighted sphere, exploiting the hierarchical structure by processing each order with its own weight, thereby achieving scale invariance without scale parameters and preserving discriminative power under unknown blur.
Timeline (last 90 days)
The paper proposes Graded Convolutional Networks, which achieve scale invariance without explicit scale parameters through a retraction operation.
Under unknown Gaussian blur, the retraction operation undergoes only a scale change, whereas scale-normalized jets produce a grading factor that cannot be eliminated by any radial normalization.
At the zero jet, retraction can separate image families that radial normalization and continuous classifiers cannot distinguish.
All Facts (3)
The paper proposes Graded Convolutional Networks, which achieve scale invariance without explicit scale parameters through a retraction operation.
50%UnverifiedUnder unknown Gaussian blur, the retraction operation undergoes only a scale change, whereas scale-normalized jets produce a grading factor that cannot be eliminated by any radial normalization.
50%UnverifiedAt the zero jet, retraction can separate image families that radial normalization and continuous classifiers cannot distinguish.
50%