MOWA: Multiple-in-One Image Warping Model

Warping Results

Results on Test Datasets. Qualitative comparison of our multiple-in-one framework MOWA to the SOTA image warping models: The red dotted lines mark the horizon, and the arrows highlight the inferior warped parts such as the irregular boundaries and distorted semantics.

Cross-Domain and Zero-Shot Evaluations. Generalization evaluation of the proposed MOWA: cross-domain evaluation (top) and zero-shot evaluation (bottom). In image retargeting results, the red dotted lines measure the stretch extent of the face or body by warping operations.

Play the Warping! Flow Visualization and Interaction

Method

Framework. To mitigate the difficulty of multi-task learning, we propose to decouple the motion estimation in both the region level (TPS control points with increasing refined numbers) and pixel level (residual flow). Additionally, a prompt learning module, guided by a lightweight point-based classifier, is designed to facilitate task-aware image warpings.

TPS Control Point. The TPS transformation stands out for its remarkable flexibility to model complex motions. It is adept at performing image warping based on two sets of region-level control points. We visualize the estimation results of control points from different levels and progressive warping results as follows.

Prompt Learning. To dynamically boost the task-aware image warping in a single model, a prompt learning module is presented in MOWA. The performance conflicts of different tasks can be relieved by the proposed prompt learning. All tasks show a similar improvement trend as the training epoch increases, and the framework achieves the unified and best warping performance on all tasks at the end of training epochs. (Recap. Task 1: stitched images, Task 2: rectified wide-angle images, Task 3: unrolling shutter images, Task 4: rotated images, Task 5: fisheye images, Task 6: portrait photos.)

More Warping Results