Skip to content
This repository has been archived by the owner on Oct 31, 2023. It is now read-only.


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time
October 8, 2019 23:46
September 25, 2019 10:26
August 20, 2019 11:17
August 20, 2019 11:17
August 20, 2019 11:17
February 19, 2020 14:50
October 2, 2019 15:44
October 2, 2019 15:43
October 2, 2019 15:41

Deep Hough Voting for 3D Object Detection in Point Clouds

Created by Charles R. Qi, Or Litany, Kaiming He and Leonidas Guibas from Facebook AI Research and Stanford University.



This repository is code release for our ICCV 2019 paper (arXiv report here).

Current 3D object detection methods are heavily influenced by 2D detectors. In order to leverage architectures in 2D detectors, they often convert 3D point clouds to regular grids (i.e., to voxel grids or to bird’s eye view images), or rely on detection in 2D images to propose 3D boxes. Few works have attempted to directly detect objects in point clouds. In this work, we return to first principles to construct a 3D detection pipeline for point cloud data and as generic as possible. However, due to the sparse nature of the data – samples from 2D manifolds in 3D space – we face a major challenge when directly predicting bounding box parameters from scene points: a 3D object centroid can be far from any surface point thus hard to regress accurately in one step. To address the challenge, we propose VoteNet, an end-to-end 3D object detection network based on a synergy of deep point set networks and Hough voting. Our model achieves state-of-the-art 3D detection on two large datasets of real 3D scans, ScanNet and SUN RGB-D with a simple design, compact model size and high efficiency. Remarkably, VoteNet outperforms previous methods by using purely geometric information without relying on color images.

In this repository, we provide VoteNet model implementation (with Pytorch) as well as data preparation, training and evaluation scripts on SUN RGB-D and ScanNet.


If you find our work useful in your research, please consider citing:

    author = {Qi, Charles R and Litany, Or and He, Kaiming and Guibas, Leonidas J},
    title = {Deep Hough Voting for 3D Object Detection in Point Clouds},
    booktitle = {Proceedings of the IEEE International Conference on Computer Vision},
    year = {2019}


Install Pytorch and Tensorflow (for TensorBoard). It is required that you have access to GPUs. Matlab is required to prepare data for SUN RGB-D. The code is tested with Ubuntu 18.04, Pytorch v1.1, TensorFlow v1.14, CUDA 10.0 and cuDNN v7.4. Note: After a code update on 2/6/2020, the code is now also compatible with Pytorch v1.2+

Compile the CUDA layers for PointNet++, which we used in the backbone network:

cd pointnet2
python install

To see if the compilation is successful, try to run python models/ to see if a forward pass works.

Install the following Python dependencies (with pip install):


Run demo

You can download pre-trained models and sample point clouds HERE. Unzip the file under the project root path (/path/to/project/demo_files) and then run:


The demo uses a pre-trained model (on SUN RGB-D) to detect objects in a point cloud from an indoor room of a table and a few chairs (from SUN RGB-D val set). You can use 3D visualization software such as the MeshLab to open the dumped file under demo_files/sunrgbd_results to see the 3D detection output. Specifically, open ***_pc.ply and ***_pred_confident_nms_bbox.ply to see the input point cloud and predicted 3D bounding boxes.

You can also run the following command to use another pretrained model on a ScanNet:

python --dataset scannet --num_point 40000

Detection results will be dumped to demo_files/scannet_results.

Training and evaluating

Data preparation

For SUN RGB-D, follow the README under the sunrgbd folder.

For ScanNet, follow the README under the scannet folder.

Train and test on SUN RGB-D

To train a new VoteNet model on SUN RGB-D data (depth images):

CUDA_VISIBLE_DEVICES=0 python --dataset sunrgbd --log_dir log_sunrgbd

You can use CUDA_VISIBLE_DEVICES=0,1,2 to specify which GPU(s) to use. Without specifying CUDA devices, the training will use all the available GPUs and train with data parallel (Note that due to I/O load, training speedup is not linear to the nubmer of GPUs used). Run python -h to see more training options (e.g. you can also set --model boxnet to train with the baseline BoxNet model). While training you can check the log_sunrgbd/log_train.txt file on its progress, or use the TensorBoard to see loss curves.

To test the trained model with its checkpoint:

python --dataset sunrgbd --checkpoint_path log_sunrgbd/checkpoint.tar --dump_dir eval_sunrgbd --cluster_sampling seed_fps --use_3d_nms --use_cls_nms --per_class_proposal

Example results will be dumped in the eval_sunrgbd folder (or any other folder you specify). You can run python -h to see the full options for evaluation. After the evaluation, you can use MeshLab to visualize the predicted votes and 3D bounding boxes (select wireframe mode to view the boxes). Final evaluation results will be printed on screen and also written in the log_eval.txt file under the dump directory. In default we evaluate with both AP@0.25 and AP@0.5 with 3D IoU on oriented boxes. A properly trained VoteNet should have around 57 mAP@0.25 and 32 mAP@0.5.

Train and test on ScanNet

To train a VoteNet model on Scannet data (fused scan):

CUDA_VISIBLE_DEVICES=0 python --dataset scannet --log_dir log_scannet --num_point 40000

To test the trained model with its checkpoint:

python --dataset scannet --checkpoint_path log_scannet/checkpoint.tar --dump_dir eval_scannet --num_point 40000 --cluster_sampling seed_fps --use_3d_nms --use_cls_nms --per_class_proposal

Example results will be dumped in the eval_scannet folder (or any other folder you specify). In default we evaluate with both AP@0.25 and AP@0.5 with 3D IoU on axis aligned boxes. A properly trained VoteNet should have around 58 mAP@0.25 and 35 mAP@0.5.

Train on your own data

[For Pro Users] If you have your own dataset with point clouds and annotated 3D bounding boxes, you can create a new dataset class and train VoteNet on your own data. To ease the proces, some tips are provided in this doc.


We want to thank Erik Wijmans for his PointNet++ implementation in Pytorch (original codebase).


votenet is relased under the MIT License. See the LICENSE file for more details.

Change log

10/20/2019: Fixed a bug of the 3D interpolation customized ops (corrected gradient computation). Re-training the model after the fix slightly improves mAP (less than 1 point).