PointRCNN is a novel 3D object detection framework that directly operates on raw point clouds to generate 3D object proposals and detect objects. The framework consists of two stages: stage-1 generates 3D proposals by segmenting the point cloud into foreground and background points, while stage-2 refines these proposals using canonical coordinates and semantic features. Unlike previous methods that rely on 2D images or voxels, PointRCNN directly generates high-quality 3D proposals from point clouds in a bottom-up manner. The stage-2 network transforms the pooled points of each proposal to canonical coordinates to learn better local spatial features, which are combined with global semantic features from stage-1 for accurate box refinement and confidence prediction. Extensive experiments on the KITTI dataset show that PointRCNN outperforms state-of-the-art methods with significant margins, achieving high recall and accuracy in 3D object detection. The framework is efficient and effective, using only point cloud data as input. The proposed bin-based loss function for 3D bounding box regression is also shown to be effective, improving the performance of the detection framework. PointRCNN's two-stage approach enables robust and accurate 3D detection, making it a promising solution for autonomous driving and other real-world applications.PointRCNN is a novel 3D object detection framework that directly operates on raw point clouds to generate 3D object proposals and detect objects. The framework consists of two stages: stage-1 generates 3D proposals by segmenting the point cloud into foreground and background points, while stage-2 refines these proposals using canonical coordinates and semantic features. Unlike previous methods that rely on 2D images or voxels, PointRCNN directly generates high-quality 3D proposals from point clouds in a bottom-up manner. The stage-2 network transforms the pooled points of each proposal to canonical coordinates to learn better local spatial features, which are combined with global semantic features from stage-1 for accurate box refinement and confidence prediction. Extensive experiments on the KITTI dataset show that PointRCNN outperforms state-of-the-art methods with significant margins, achieving high recall and accuracy in 3D object detection. The framework is efficient and effective, using only point cloud data as input. The proposed bin-based loss function for 3D bounding box regression is also shown to be effective, improving the performance of the detection framework. PointRCNN's two-stage approach enables robust and accurate 3D detection, making it a promising solution for autonomous driving and other real-world applications.