FGR: Frustum-Aware Geometric Reasoning for Weakly Supervised 3D Vehicle Detection

MAP-Gen: An Automated 3D-Box Annotation Flow with Multimodal Attention Point Generator

“…Later, Frustum PointNet [14] generates 2D proposals on the image and then uses cascading PointNet [22] to predict 3D boxes. Similarly, FGR [13] extracts preliminary frustum sub clouds from which 3D annotations are generated. On the other hand, PI-RCNN [26] and PointPainting [17] try to directly fuse point cloud features with the image information.…”

“…Several previous studies deploy weak annotations, such as center-clicks [7], [8], extreme-points [9], [10], polygonpoints [11], [12], 2D bounding boxes [13], [14], or 2D segmentation masks [15], [16]. To facilitate 3D object detection with lower annotation costs, a few existing methods enhance the weak labels into a stronger form, for example, from 2D (c) Image-guided interpolation.…”

“…bounding boxes to 3D boxes [13]. After that, object detectors can be trained under weak supervision (i.e., 2D boxes), given 3D pseudo labels generated by these autolabelers.…”

“…To handle such sparsity problem, some existing methods [13], [14], [17] leverage RGB images to coarsely locate or distinguish objects in the sparse point clouds, as shown in Fig. 1a and Fig.…”

“…traffic light pole) in images than in LiDAR point clouds [17]. However, almost all existing multimodal autolabelers [7], [13], [15], [18], [19] take images as auxiliary, either cropping/zooming to particular regions of the point cloud or fusing point features with image features. Although they are the current state-of-the-art methods, we argue this approach does not fully address the sparsity problem because the cloud sizes remain unchanged.…”