Painting Outside as Inside: Edge Guided Image Outpainting via Bidirectional Rearrangement with Progressive Step Learning

In-N-Out: Towards Good Initialization for Inpainting and Outpainting

“…Specifically, our method (In-N-Out) lets a model first experience the opposite task (e.g., inpainting), which is transferred to the target task (e.g., outpainting). We show by the experiment (Table 1c) that our simple transfer strategy gives an equivalent or a larger amount of performance gain to the task-specific (i.e., horizontal extrapolation) method [16], which also exploits the cross-task relation. Compared to common practice for the inpainting or outpainting, where pretraining on the target task is performed, our method gives no overhead, since it substitutes the pretraining task with the opposite task without additional overhead.…”

“…We simplify and show the relationship in the form of transfer learning. We empirically show that this simple approach can have an equivalent or better performance than the task-specific method [16].…”

“…Regarding that both tasks do not require a specific label, i.e., self-supervised, exploiting the opposite task is feasible and desirable. It has been shown that exploring across the tasks is beneficial by solving outpainting with inpainting [16], or learning cycle-consistency between both tasks [14]. We step further by simplifying the cross-task learning and generalize it to both tasks with transfer learning, which is a well-known general strategy in deep learning [40].…”

“…Intuition for transferability is based on the complementary property between inpainting and outpainting. Inpainting and outpainting have been taken to account together [14,16]. We simplify and show the relationship in the form of transfer learning.…”

“…In-N-Out is motivated by the complementary relationship between image inpainting and image outpainting tasks. For instance, changing an outpainting task to an inpainting task [16] and cycle consistency between both tasks [14] are introduced. Inspired by this, we propose our method, In-N-Out, which is a simple and general Though our method, In-N-Out, shows worse performance in the training stage, In-N-Out converges more fastly and shows better performance in the end, where each method learns each target task by fine-tuning.…”