DRACO: Weakly Supervised Dense Reconstruction And Canonicalization of Objects

We present DRACO, a method for Dense Reconstruction And Canonicalization of\nObject shape from one or more RGB images. Canonical shape reconstruction,\nestimating 3D object shape in a coordinate space canonicalized for scale,\nrotation, and translation parameters, is an emerging paradigm that holds\npromise for a multitude of robotic applications. Prior approaches either rely\non painstakingly gathered dense 3D supervision, or produce only sparse\ncanonical representations, limiting real-world applicability. DRACO performs\ndense canonicalization using only weak supervision in the form of camera poses\nand semantic keypoints at train time. During inference, DRACO predicts dense\nobject-centric depth maps in a canonical coordinate-space, solely using one or\nmore RGB images of an object. Extensive experiments on canonical shape\nreconstruction and pose estimation show that DRACO is competitive or superior\nto fully-supervised methods.\n