Abstract
Reconstructing 3D objects is an important computer vision task that has wideapplication in AR/VR. Deep learning algorithm developed for this task usuallyrelies on an unrealistic synthetic dataset, such as ShapeNet and Things3D. Onthe other hand, existing real-captured object-centric datasets usually do nothave enough annotation to enable supervised training or reliable evaluation. Inthis technical report, we present a photo-realistic object-centric datasetHM3D-ABO. It is constructed by composing realistic indoor scene and realisticobject. For each configuration, we provide multi-view RGB observations, awater-tight mesh model for the object, ground truth depth map and object mask.The proposed dataset could also be useful for tasks such as camera poseestimation and novel-view synthesis. The dataset generation code is released athttps://github.com/zhenpeiyang/HM3D-ABO.