In many robotics and VR/AR applications, 3D-videos are readily-available sources of input (a continuous sequence of depth images, or LIDAR scans). However, those 3D-videos are processed frame-by-frame either through 2D convnets or 3D perception algorithms. In this work, we propose 4-dimensional convolutional neural networks for spatio-temporal perception that can directly process such 3D-videos using high-dimensional convolutions. For this, we adopt sparse tensors and propose the generalized sparse convolution that encompasses all discrete convolutions. To implement the generalized sparse convolution, we create an open-source auto-differentiation library for sparse tensors that provides extensive functions for high-dimensional convolutional neural networks. We create 4D spatio-temporal convolutional neural networks using the library and validate them on various 3D semantic segmentation benchmarks and proposed 4D datasets for 3D-video perception. To overcome challenges in the 4D space, we propose the hybrid kernel, a special case of the generalized sparse convolution, and the trilateral-stationary conditional random field that enforces spatio-temporal consistency in the 7D space-time-chroma space. Experimentally, we show that convolutional neural networks with only generalized 3D sparse convolutions can outperform 2D or 2D-3D hybrid methods by a large margin. Also, we show that on 3D-videos, 4D spatio-temporal convolutional neural networks are robust to noise, outperform 3D convolutional neural networks and are faster than the 3D counterpart in some cases.
We implement MinkUNet with [TorchSparse](https://github.com/mit-han-lab/torchsparse) / [Minkowski Engine](https://github.com/NVIDIA/MinkowskiEngine) / [Spconv](https://github.com/traveller59/spconv) backend and provide the result and checkpoints on SemanticKITTI datasets.
## Results and models
### SemanticKITTI
| Method | Backend | Lr schd | Amp | Laser-Polar Mix | Mem (GB) | Training Time (hours) | FPS | mIoU | Download |
**Note:** We follow the implementation in SPVNAS original [repo](https://github.com/mit-han-lab/spvnas) and W16\\W20\\W32 indicates different number of channels.
**Note:** Due to TorchSparse backend, the model performance is unstable with TorchSparse backend and may fluctuate by about 1.5 mIoU for different random seeds.
**Note:** Referring to [PCSeg](https://github.com/PJLab-ADG/PCSeg), MinkUNet34v2 is modified based on MinkUNet34.
**Note\*:** Training Time and FPS are measured on NVIDIA A100. The versions of Torchsparse, Minkowski Engine and Spconv are 0.5.4, 1.4.0 and 2.3.6 respectively. Since spconv 2.3.6 has a bug with fp16 on in the inference stage, the actual FPS measurement using fp32.
# Objects are Different: Flexible Monocular 3D Object Detection
> [Objects are Different: Flexible Monocular 3D Object Detection](https://arxiv.org/abs/2104.02323)
<!-- [ALGORITHM] -->
## Abstract
The precise localization of 3D objects from a single image without depth information is a highly challenging problem. Most existing methods adopt the same approach for all objects regardless of their diverse distributions, leading to limited performance for truncated objects. In this paper, we propose a flexible framework for monocular 3D object detection which explicitly decouples the truncated objects and adaptively combines multiple approaches for object depth estimation. Specifically, we decouple the edge of the feature map for predicting long-tail truncated objects so that the optimization of normal objects is not influenced. Furthermore, we formulate the object depth estimation as an uncertainty-guided ensemble of directly regressed object depth and solved depths from different groups of keypoints. Experiments demonstrate that our method outperforms the state-of-the-art method by relatively 27% for the moderate level and 30% for the hard level in the test set of KITTI benchmark while maintaining real-time efficiency.
Note: mAP represents Car moderate 3D strict AP11 / AP40 results. Because of the limited data for pedestrians and cyclists, the detection performance for these two classes is usually unstable. Therefore, we only list car detection results here. In addition, the AP11 result may fluctuate in a larger range (~1 AP), so AP40 is a more recommended metric for reference due to its much better stability.
## Citation
```latex
@InProceedings{MonoFlex,
author = {Zhang, Yunpeng and Lu, Jiwen and Zhou, Jie},
title = {Objects Are Different: Flexible Monocular 3D Object Detection},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
In this technical report, we present our solution, dubbed MV-FCOS3D++, for the Camera-Only 3D Detection track in Waymo Open Dataset Challenge 2022. For multi-view camera-only 3D detection, methods based on bird-eye-view or 3D geometric representations can leverage the stereo cues from overlapped regions between adjacent views and directly perform 3D detection without hand-crafted post-processing. However, it lacks direct semantic supervision for 2D backbones, which can be complemented by pretraining simple monocular-based detectors. Our solution is a multi-view framework for 4D detection following this paradigm. It is built upon a simple monocular detector FCOS3D++, pretrained only with object annotations of Waymo, and converts multi-view features to a 3D grid space to detect 3D objects thereon. A dual-path neck for single-frame understanding and temporal stereo matching is devised to incorporate multi-frame information. Our method finally achieves 49.75% mAPL with a single model and wins 2nd place in the WOD challenge, without any LiDAR-based depth supervision during training. The code will be released at [this https URL](https://github.com/Tai-Wang/Depth-from-Motion).
Regrettably, we are unable to provide the pre-trained model weights due to [Waymo Dataset License Agreement](https://waymo.com/open/terms/), so we only provide the training logs as shown above.
## Citation
```latex
@article{wang2022mvfcos3d++,
title={{MV-FCOS3D++: Multi-View} Camera-Only 4D Object Detection with Pretrained Monocular Backbones},
author={Wang, Tai and Lian, Qing and Zhu, Chenming and Zhu, Xinge and Zhang, Wenwei},
# MVX-Net: Multimodal VoxelNet for 3D Object Detection
> [MVX-Net: Multimodal VoxelNet for 3D Object Detection](https://arxiv.org/abs/1904.01649)
<!-- [ALGORITHM] -->
## Abstract
Many recent works on 3D object detection have focused on designing neural network architectures that can consume point cloud data. While these approaches demonstrate encouraging performance, they are typically based on a single modality and are unable to leverage information from other modalities, such as a camera. Although a few approaches fuse data from different modalities, these methods either use a complicated pipeline to process the modalities sequentially, or perform late-fusion and are unable to learn interaction between different modalities at early stages. In this work, we present PointFusion and VoxelFusion: two simple yet effective early-fusion approaches to combine the RGB and point cloud modalities, by leveraging the recently introduced VoxelNet architecture. Evaluation on the KITTI dataset demonstrates significant improvements in performance over approaches which only use point cloud data. Furthermore, the proposed method provides results competitive with the state-of-the-art multimodal algorithms, achieving top-2 ranking in five of the six bird's eye view and 3D detection categories on the KITTI benchmark, by using a simple single stage network.
