OverallNet: Scale-Arbitrary Lightweight SR Model
for handling 360° Panoramic Images
SIGGRAPH ASIA Poster, 2024
Dongsik Yoon, Jongeun Kim, Seonggeun Song, Yejin Lee, Gunhee Lee
HDC LABS
A summary of the method for processing panoramics with the SR technique. Please zoom in to check more details.
Abstract
Estimating room layouts from 360o panoramic images is an essen- tial task in computer vision, enhancing 3D scene understanding from 2D images. To achieve and handle high-quality panoramas necessitates a Super-Resolution (SR) technique that is both light and capable of arbitrary-scale SR, optimized for efficient inference.
In this study, we propose a simple model that employs a modular arbitrary-scale technique. Additionally, our model incorporates quantization to maximize efficiency during user inference, making it well-suited for processing high-quality panoramic images.
Proposed Methods
We propose a novel architecture that incorporates the scale-aware module of ArbSR. To further enhance SR sharpness and realism, we introduce a feature extractor built with a block called the scale-aware dense group SADG.
Unlike the existing OSM, our proposed architecture includes a scale-aware upsampling layer, facilitating scale-arbitrary SR at various factors. Moreover, recent research suggests that the Swish activation function outperforms the ReLU activation function in SR tasks, we replaced all activation functions in the SADG block with SiLU layers.
Qualitative Comparison
We compare the SR quality with the existing OverNet method, each trained on three different scale types. For this work, we conduct validations on the Set14 dataset. Figure presents a comparison between images produced by our proposed method and those generated by OverNet. It is evident that OverNet shows considerable degradation in performance when applied to scales not covered in its training dataset.