- [pdf] [code]
Meta-Det3D: Learn to Learn Few-Shot 3D Object Detection
This paper addresses the problem of few-shot indoor 3D object detection by proposing a meta-learning-based framework that only relies on a few labeled samples from novel classes for training. Our model has two major components: a 3D meta-detector and a 3D object detector. Given a query 3D point cloud and a few support samples, the 3D meta-detector is trained over different 3D detection tasks to learn task distributions for different object classes and dynamically adapt the 3D object detector to complete a specific detection task. The 3D object detector takes task-specific information as input and produces 3D object detection results for the query point cloud. Specifically, the 3D object detector first extracts object candidates and their features from the query point cloud using a point feature learning network. Then, a class-specific re-weighting module generates class-specific re-weighting vectors from the support samples to characterize the task information, one for each distinct object class. Each re-weighting vector performs channel-wise attention to the candidate features to re-calibrate the query object features, adapting them to detect objects of the same classes. Finally, the adapted features are fed into a detection head to predict classification scores and bounding boxes for novel objects in the query point cloud. Several experiments on two 3D object detection benchmark datasets demonstrate that our proposed method acquired the ability to detect 3D objects in the few-shot setting.