Improving 3D Vulnerable Road User Detection With Point Augmentation

Point clouds have been a popular representation to describe 3D environments for autonomous driving applications. Despite accurate depth information, sparsity of points results in difficulties extracting sufficient features from vulnerable objects small sizes. One solution is leveraging self-attention networks build long-range connections between similar objects. Another method using generative models estimate the complete shape Both approaches introduce large memory consumption and extra complexity while geometric characteristics are overlooked. To overcome this problem, paper proposes Augmentation (PA)-RCNN, focusing on object detection by generating efficient complementary without trainable parameters. Specifically, sampled with guidance proposals encoded through grid-based feature aggregation produce localised voxel properties. Such attributes fed pooling module aid fictional points, which transformed considering symmetry. Experimental Waymo Open Dataset KITTI dataset show superior advantage distant comparison existing state-of-the-art methods.