synScan - A Large-Scale Dataset for Instance Level Recognition on Partial Scan Data

Abstract

Devices supporting depth-sensing technologies become more and more available, making it easier to access geometry-data driven services like 3D model or scene reconstruction. Utilizing these depth sensors, very large datasets have been created to enable deep learning for object detection and depth upsampling. We want to tackle the task of instance level recognition (ILR), where 3D scans of objects can be searched against a database of CAD models based on embeddings of their geometry. The distinctive property of this retrieval task is the existence of only a single corresponding database shape for each query. To the best of our knowledge all the existing datasets either lack in providing the exact CAD model correspondences or lack in scale and a variety of object categories. Therefore, we introduce synScan, a large-scale dataset synthetically generated via physically plausible domain randomization (PPDR) of 3D scenes and object-centric scan trajectories with the goal to mimic real-world object scan scenarios with a variety of incomplete views and occlusions. We provide approximately 39,000 randomly sampled scenes, made from 9,400 different shapes with semantic per-point annotation. We train and test different ILR algorithms (e.g. PointNetVLAD, MinkLoc3Dv2) designed for place-recognition in self-driving cars on the dataset and validate our results on a smaller real-world dataset. Utilizing a rather simple data generation pipeline, we can show that deep learning methods trained on our synthetic dataset can successfully adapt to real-world scan data. In this manner, synScan helps to overcome the lack of labeled training data.