MultiPanoWise: holistic deep architecture for multi-task dense prediction from a single panoramic image

We present a novel holistic deep-learning approach for multi-task learning from a single indoor panoramic image. Our framework, named MultiPanoWise, extends vision transformers to jointly infer multiple pixel-wise signals, such as depth, normals, and semantic segmentation, as well as signals from intrinsic decomposition, such as reflectance and shading. Our solution leverages a specific architecture combining a transformer-based encoder-decoder with multiple heads, by introducing, in particular, a novel context adjustment approach, to enforce knowledge distillation between the various signals. Moreover, at training time we introduce a hybrid loss scalarization method based on an augmented Chebychev/hypervolume scheme. We illustrate the capabilities of the proposed architecture on public-domain synthetic and real-world datasets. We demonstrate performance improvements with respect to the most recent methods specifically designed for single tasks, like, for example, individual depth estimation or semantic segmentation. To our knowledge, this is the first architecture capable of achieving state-of-the-art performance on the joint extraction of heterogeneous signals from single indoor omnidirectional images.