3DVG-Transformer
This repository is for the ICCV 2021 paper "3DVG-Transformer: Relation Modeling for Visual Grounding on Point Clouds"
Our method "3DVG-Transformer+" is the 1st method on the ScanRefer benchmark (2021/3 - 2021/11) and is the winner of the CVPR2021 1st Workshop on Language for 3D Scenes
Introduction
Visual grounding on 3D point clouds is an emerging vision and language task that benefits various applications in understanding the 3D visual world. By formulating this task as a grounding-by-detection problem, lots of recent works focus on how to exploit more powerful detectors and comprehensive language features, but (1) how to model complex relations for generating context-aware object proposals and (2) how to leverage proposal relations to distinguish the true target object from similar proposals are not fully studied yet. Inspired by the well-known transformer architecture, we propose a relation-aware visual grounding method on 3D point clouds, named as 3DVG-Transformer, to fully utilize the contextual clues for relation-enhanced proposal generation and cross-modal proposal disambiguation, relation-aware proposal generation and cross-modal feature fusion, which are enabled by a newly designed coordinate-guided contextual aggregation (CCA) module in the object proposal generation stage, and a multiplex attention (MA) module in the cross-modal feature fusion stage. With the aid of two proposed feature augmentation strategies to alleviate overfitting, we validate that our 3DVG-Transformer outperforms the state-of-the-art methods by a large margin, on two point cloud-based visual grounding datasets, ScanRefer and Nr3D/Sr3D from ReferIt3D, especially for complex scenarios containing multiple objects of the same category.
Dataset & Setup
Data preparation
This codebase is built based on the initial ScanRefer codebase. Please refer to ScanRefer for more data preprocessing details.
- Download the ScanRefer dataset and unzip it under
data/
. - Downloadand the preprocessed GLoVE embeddings (~990MB) and put them under
data/
. - Download the ScanNetV2 dataset and put (or link)
scans/
under (or to)data/scannet/scans/
(Please follow the ScanNet Instructions for downloading the ScanNet dataset).
After this step, there should be folders containing the ScanNet scene data under the
data/scannet/scans/
with names likescene0000_00
- Pre-process ScanNet data. A folder named
scannet_data/
will be generated underdata/scannet/
after running the following command. Roughly 3.8GB free space is needed for this step:
cd data/scannet/
python batch_load_scannet_data.py
After this step, you can check if the processed scene data is valid by running:
python visualize.py --scene_id scene0000_00
- (Optional) Pre-process the multiview features from ENet.
-
Download: Download the ENet multiview features (~36GB, hdf5 database) and put it under
data/scannet/scannet_data/
-
Projection:
a. Download the ENet pretrained weights (1.4MB) and put it under
data/
b. Download and decompress the extracted ScanNet frames (~13GB). c. Change the data paths inlib/config.py
marked with TODO accordingly. d. Project ENet features from ScanNet frames to point clouds (~36GB, hdf5 database).
python script/project_multiview_features.py --maxpool
Setup
The code is tested on Ubuntu 16.04 LTS & 18.04 LTS with PyTorch 1.2.0 CUDA 10.0 installed.
Please refer to the initial ScanRefer for pointnet2 packages for the newer version (>=1.3.0) of PyTorch.
You could use other PointNet++ implementations for the lower version (<=1.2.0) of PyTorch.
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch
Install the necessary packages listed out in requirements.txt
:
pip install -r requirements.txt
After all packages are properly installed, please run the following commands to compile the CUDA modules for the PointNet++ backbone:
cd lib/pointnet2
python setup.py install
Before moving on to the next step, please don't forget to set the project root path to the CONF.PATH.BASE
in lib/config.py
.
Usage
Training
To train the 3DVG-Transformer model with multiview features:
python scripts/ScanRefer_train.py --use_multiview --use_normal --batch_size 8 --epoch 200 --lr 0.002 --coslr --tag 3dvg-trans+
settings: XYZ: --use_normal XYZ+RGB: --use_color --use_normal XYZ+Multiview: --use_multiview --use_normal
For more training options (like using preprocessed multiview features), please run scripts/train.py -h
.
Evaluation
To evaluate the trained models, please find the folder under outputs/
and run:
python scripts/ScanRefer_eval.py --folder <folder_name> --reference --use_multiview --no_nms --force --repeat 5 --lang_num_max 1
Note that the flags must match the ones set before training. The training information is stored in outputs/<folder_name>/info.json
Note that the results generated by ScanRefer_eval.py may be slightly lower than the test results during training. The main reason is that the results of model testing fluctuate, while the maximum value is reported during training, and we do not use a fixed test seed.
Benchmark Challenge
Note that every user is allowed to submit the test set results of each method only twice, and the ScanRefer benchmark blocks update the test set results of a method for two weeks after a test set submission.
After finishing training the model, please download the benchmark data and put the unzipped ScanRefer_filtered_test.json
under data/
. Then, you can run the following script the generate predictions:
python benchmark/predict.py --folder <folder_name> --use_color
Note that the flags must match the ones set before training. The training information is stored in outputs/<folder_name>/info.json
. The generated predictions are stored in outputs/<folder_name>/pred.json
. For submitting the predictions, please compress the pred.json
as a .zip or .7z file and follow the instructions to upload your results.
Visualization
To predict the localization results predicted by the trained ScanRefer model in a specific scene, please find the corresponding folder under outputs/
with the current timestamp and run:
python scripts/visualize.py --folder <folder_name> --scene_id <scene_id> --use_color
Note that the flags must match the ones set before training. The training information is stored in outputs/<folder_name>/info.json
. The output .ply
files will be stored under outputs/<folder_name>/vis/<scene_id>/
In our next version, the heatmap visualization code will be open-sourced in the 3DJCG (CVPR2022, Oral) codebase.
The generated .ply or .obj files could be visualized in software such as MeshLab.
Results
settings: 3D Only (XYZ+RGB): --use_color --use_normal 2D+3D (XYZ+Multiview): --use_multiview --use_normal
Validation Set | Unique | Unique | Multiple | Multiple | Overall | Overall | ||
---|---|---|---|---|---|---|---|---|
Methods | Publication | Modality | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] |
SCRC | CVPR16 | 2D | 24.03 | 9.22 | 17.77 | 5.97 | 18.70 | 6.45 |
One-Stage | ICCV19 | 2D | 29.32 | 22.82 | 18.72 | 6.49 | 20.38 | 9.04 |
ScanRefer | ECCV2020 | 3D | 67.64 | 46.19 | 32.06 | 21.26 | 38.97 | 26.10 |
TGNN | AAAI2021 | 3D | 68.61 | 56.80 | 29.84 | 23.18 | 37.37 | 29.70 |
InstanceRefer | ICCV2021 | 3D | 77.45 | 66.83 | 31.27 | 24.77 | 40.23 | 32.93 |
SAT | ICCV2021 | 3D | 73.21 | 50.83 | 37.64 | 25.16 | 44.54 | 30.14 |
3DVG-Transformer (ours) | ICCV2021 | 3D | 77.16 | 58.47 | 38.38 | 28.70 | 45.90 | 34.47 |
BEAUTY-DETR | - | 3D | - | - | - | - | 46.40 | - |
3DJCG | CVPR2022 | 3D | 78.75 | 61.30 | 40.13 | 30.08 | 47.62 | 36.14 |
3D-SPS | CVPR2022 | 3D | 81.63 | 64.77 | 39.48 | 29.61 | 47.65 | 36.43 |
ScanRefer | ECCV2020 | 2D + 3D | 76.33 | 53.51 | 32.73 | 21.11 | 41.19 | 27.40 |
TGNN | AAAI2021 | 2D + 3D | 68.61 | 56.80 | 29.84 | 23.18 | 37.37 | 29.70 |
InstanceRefer | ICCV2021 | 2D + 3D | 75.72 | 64.66 | 29.41 | 22.99 | 38.40 | 31.08 |
3DVG-Transformer (Ours) | ICCV2021 | 2D + 3D | 81.93 | 60.64 | 39.30 | 28.42 | 47.57 | 34.67 |
3DVG-Transformer+(Ours, this codebase) | - | 2D + 3D | 83.25 | 61.95 | 41.20 | 30.29 | 49.36 | 36.43 |
MVT-3DVG | CVPR2022 | 2D + 3D | 77.67 | 66.45 | 31.92 | 25.26 | 40.80 | 33.26 |
3DJCG | CVPR2022 | 2D + 3D | 83.47 | 64.34 | 41.39 | 30.82 | 49.56 | 37.33 |
3D-SPS | CVPR2022 | 2D + 3D | 84.12 | 66.72 | 40.32 | 29.82 | 48.82 | 36.98 |
Online Benchmark | Unique | Unique | Multiple | Multiple | Overall | Overall | |
---|---|---|---|---|---|---|---|
Methods | Modality | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] |
ScanRefer | 2D + 3D | 68.59 | 43.53 | 34.88 | 20.97 | 42.44 | 26.03 |
TGNN | 2D + 3D | 68.34 | 58.94 | /33.12 | 25.26 | 41.02 | 32.81 |
InstanceRefer | 2D + 3D | 77.82 | 66.69 | 34.57 | 26.88 | 44.27 | 35.80 |
3DVG-Transformer (Ours) | 2D + 3D | 75.76 | 55.15 | 42.24 | 29.33 | 49.76 | 35.12 |
3DVG-Transformer+(Ours) | 2D + 3D | 77.33 | 57.87 | 43.70 | 31.02 | 51.24 | 37.04 |
Changelog
2022/04: Update Readme.md.
2022/04: Release the codes of 3DVG-Transformer.
2021/07: 3DVG-Transformer is accepted at ICCV 2021.
2021/06: 3DVG-Transformer+ won the ScanRefer Challenge in the CVPR2021 1st Workshop on Language for 3D Scenes.
2021/04: 3DVG-Transformer+ achieves 1st place in ScanRefer Leaderboard.
Citation
If you use the codes in your work, please kindly cite our work 3DVG-Transformer and the original ScanRefer paper:
@inproceedings{zhao2021_3DVG_Transformer,
title={{3DVG-Transformer}: Relation modeling for visual grounding on point clouds},
author={Zhao, Lichen and Cai, Daigang and Sheng, Lu and Xu, Dong},
booktitle={ICCV},
pages={2928--2937},
year={2021}
}
@article{chen2020scanrefer,
title={{ScanRefer}: 3D Object Localization in RGB-D Scans using Natural Language},
author={Chen, Dave Zhenyu and Chang, Angel X and Nie{\ss}ner, Matthias},
pages={202--221},
journal={ECCV},
year={2020}
}
Acknowledgement
We would like to thank facebookresearch/votenet for the 3D object detection codebase and erikwijmans/Pointnet2_PyTorch for the CUDA accelerated PointNet++ implementation.
For further acceleration, you could use KD-Tree to accelerate the PointNet++ process.
License
This repository is released under MIT License (see LICENSE file for details).