3D Box Proposals From a Single Monocular Image of an Indoor Scene

• DOI: 10.1609/aaai.v32i1.12314
• Publication Date: 2022-06-24T21:10:27Z
• Authors: Wei Zhuo; Mathieu Salzmann; Xuming He; Miaomiao Liu
• Keywords: Minimum bounding box; Depth map; Monocular; Bounding overwatch; Representation

Modern object detection methods typically rely on bounding box proposals as input. While initially popularized in the 2D case, this idea has received increasing attention for 3D boxes. Nevertheless, existing proposal techniques all assume having access to depth input, which is unfortunately not always available practice. In paper, we therefore introduce an approach generating from a single monocular RGB image. To end, develop integrated, fully differentiable framework that inherently predicts map, extracts volumetric scene representation and generates proposals. At core of our lies novel residual, truncated signed distance function module, which, accounting relatively low accuracy predicted scene. Our experiments standard NYUv2 dataset demonstrate lets us generate high-quality it outperforms two-stage technique consisting successively performing state-of-the-art prediction depth-based generation.