DeepSIM: Image Shape Manipulation from a Single Augmented Training Sample

Yael Vinker^*, Eliahu Horwitz^*, Nir Zabari , Yedid Hoshen

School of Computer Science and Engineering
The Hebrew University of Jerusalem, Israel
ICCV 2021 (Oral)
^*Indicates Equal Contribution

Paper Supplementary Code arXiv

DeepSIM: Given a single real training image (b) and a corresponding primitive representation (a), our model learns to map between the primitive (a) to the target image (b). At inference, the original primitive (a) is manipulated by the user. Then, the manipulated primitive is passed through the network which outputs a corresponding manipulated image (e) in the real image domain.

Abstract

We present DeepSIM, a generative model for conditional image manipulation based on a single image. We find that extensive augmentation is key for enabling single image training, and incorporate the use of thin-plate-spline (TPS) as an effective augmentation. Our network learns to map between a primitive representation of the image to the image itself. The choice of a primitive representation has an impact on the ease and expressiveness of the manipulations and can be automatic (e.g. edges), manual (e.g. segmentation) or hybrid such as edges on top of segmentations. At manipulation time, our generator allows for making complex image changes by modifying the primitive input representation and mapping it through the network. Our method is shown to achieve remarkable performance on image manipulation tasks.

DeepSIM was trained on a single training pair, shown to the left of each sample. First row "face" output- (left) flipping eyebrows, (right) lifting nose. Second row "dog" output- changing shape of dog's hat, removing ribbon, and making face longer. Second row "car" output- (top) adding wheel, (bottom) conversion to sports car.

Top row - primitive images. Left - original pair used for training. Center- switching the positions between the two rightmost cars. Right- removing the leftmost car and inpainting the background.

The leftmost column shows the source image, then each column demonstrate the result of our model when trained on the specified primitive. We manipulated the image primitives, adding a right eye, changing the point of view and shortening the beak. Our results are presented next to each manipulated primitive. The combined primitive performed best on high-level changes (e.g. the eye), and low-level changes (e.g. the background).

On the left is the training image pair, in the middle are the manipulated primitives and on the right are the manipulated outputs- left to right: dress length, strapless, wrap around the neck.

ICCV Presentation

Single Image Animation

ICCV Poster

BibTeX

@InProceedings{Vinker_2021_ICCV,
        author    = {Vinker, Yael and Horwitz, Eliahu and Zabari, Nir and Hoshen, Yedid},
        title     = {Image Shape Manipulation From a Single Augmented Training Sample},
        booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
        month     = {October},
        year      = {2021},
        pages     = {13769-13778}
      }

The single image animation primitives were created by Jonathan Reich jonathanreichandco@gmail.com