Votenet

VoteNet operates directly on raw point clouds without converting to voxels or 2D images, preserving 3D structure and simplifying the detection pipeline, as emphasized in the README's introduction.

Achieves high mAP scores on real-world datasets, with 57 mAP@0.25 and 32 mAP@0.5 on SUN RGB-D, outperforming previous methods using purely geometric information, as documented in the evaluation results.

Features a simple architecture with compact model size and high computational efficiency, making it suitable for deployment in resource-constrained scenarios, as highlighted in the key features.

The network is fully trainable from input point clouds to output bounding boxes, enabling seamless integration into deep learning workflows without manual feature engineering, as stated in the README.