ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection

News:

🔥 August, 2022. ImVoxelNet for SUN RGB-D is now supported in mmdetection3d.
🔥 October, 2021. Our paper is accepted at WACV 2022. We simplify 3d neck to make indoor models much faster and accurate. For example, this improves ScanNet mAP by more than 2%. Please find updated configs in configs/imvoxelnet/*_fast.py and models.
🔥 August, 2021. We adapt center sampling for indoor detection. For example, this improves ScanNet mAP by more than 5%. Please find updated configs in configs/imvoxelnet/*_top27.py and models.
🔥 July, 2021. We update ScanNet image preprocessing both here and in mmdetection3d.
🔥 June, 2021. ImVoxelNet for KITTI is now supported in mmdetection3d.

This repository contains implementation of the monocular/multi-view 3D object detector ImVoxelNet, introduced in our paper:

ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection
Danila Rukhovich, Anna Vorontsova, Anton Konushin
Samsung Research
https://arxiv.org/abs/2106.01178

Installation

For convenience, we provide a Dockerfile. Alternatively, you can install all required packages manually.

This implementation is based on mmdetection3d framework. Please refer to the original installation guide install.md, replacing open-mmlab/mmdetection3d with saic-vul/imvoxelnet. Also, rotated_iou should be installed with these 4 commands.

Most of the ImVoxelNet-related code locates in the following files: detectors/imvoxelnet.py, necks/imvoxelnet.py, dense_heads/imvoxel_head.py, pipelines/multi_view.py.

Datasets

We support three benchmarks based on the SUN RGB-D dataset.

For the VoteNet benchmark with 10 object categories, you should follow the instructions in sunrgbd.
For the PerspectiveNet benchmark with 30 object categories, the same instructions can be applied; you only need to set dataset argument to sunrgbd_monocular when running create_data.py.
The Total3DUnderstanding benchmark implies detecting objects of 37 categories along with camera pose and room layout estimation. Download the preprocessed data as train.json and val.json and put it to ./data/sunrgbd. Then run:
```
python tools/data_converter/sunrgbd_total.py
```

For ScanNet please follow instructions in scannet. For KITTI and nuScenes, please follow instructions in getting_started.md.

Getting Started

Please see getting_started.md for basic usage examples.

Training

To start training, run dist_train with ImVoxelNet configs:

bash tools/dist_train.sh configs/imvoxelnet/imvoxelnet_kitti.py 8

Testing

Test pre-trained model using dist_test with ImVoxelNet configs:

bash tools/dist_test.sh configs/imvoxelnet/imvoxelnet_kitti.py \
    work_dirs/imvoxelnet_kitti/latest.pth 8 --eval mAP

Visualization

Visualizations can be created with test script. For better visualizations, you may set score_thr in configs to 0.15 or more:

python tools/test.py configs/imvoxelnet/imvoxelnet_kitti.py \
    work_dirs/imvoxelnet_kitti/latest.pth --show \
    --show-dir work_dirs/imvoxelnet_kitti

Models

v2 adds center sampling for indoor scenario. v3 simplifies 3d neck for indoor scenario. Differences are discussed in v2 and v3 preprints.

Dataset	Object Classes	Version	Download
SUN RGB-D	37 from Total3dUnderstanding	v1 \| mAP@0.15: 41.5 v2 \| mAP@0.15: 42.7 v3 \| mAP@0.15: 43.7	model \| log \| config model \| log \| config model \| log \| config
SUN RGB-D	30 from PerspectiveNet	v1 \| mAP@0.15: 44.9 v2 \| mAP@0.15: 47.2 v3 \| mAP@0.15: 48.7	model \| log \| config model \| log \| config model \| log \| config
SUN RGB-D	10 from VoteNet	v1 \| mAP@0.25: 38.8 v2 \| mAP@0.25: 39.4 v3 \| mAP@0.25: 40.7	model \| log \| config model \| log \| config model \| log \| config
ScanNet	18 from VoteNet	v1 \| mAP@0.25: 40.6 v2 \| mAP@0.25: 45.7 v3 \| mAP@0.25: 48.1	model \| log \| config model \| log \| config model \| log \| config
KITTI	Car	v1 \| AP@0.7: 17.8	model \| log \| config
nuScenes	Car	v1 \| AP: 51.8	model \| log \| config

Example Detections

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{rukhovich2022imvoxelnet,
  title={Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection},
  author={Rukhovich, Danila and Vorontsova, Anna and Konushin, Anton},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={2397--2406},
  year={2022}
}