diff --git a/README.md b/README.md index f4dd01fd5c..97607656ce 100644 --- a/README.md +++ b/README.md @@ -251,6 +251,7 @@ Results and models are available in the [model zoo](docs/en/model_zoo.md). | SA-SSD | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | | FCAF3D | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | | PV-RCNN | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | +| SPVCNN | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | **Note:** All the about **300+ models, methods of 40+ papers** in 2D detection supported by [MMDetection](https://github.com/open-mmlab/mmdetection/blob/3.x/docs/en/model_zoo.md) can be trained or used in this codebase. diff --git a/README_zh-CN.md b/README_zh-CN.md index 9ec26b9ef3..31c13c0b81 100644 --- a/README_zh-CN.md +++ b/README_zh-CN.md @@ -232,6 +232,7 @@ MMDetection3D 是一个基于 PyTorch 的目标检测开源工具箱，下一代 | SA-SSD | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | | FCAF3D | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | | PV-RCNN | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | +| SPVCNN | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | **注意：**[MMDetection](https://github.com/open-mmlab/mmdetection/blob/3.x/docs/zh_cn/model_zoo.md) 支持的基于 2D 检测的 **300+ 个模型，40+ 的论文算法**在 MMDetection3D 中都可以被训练或使用。 diff --git a/configs/spvcnn/README.md b/configs/spvcnn/README.md new file mode 100644 index 0000000000..290cd2aac0 --- /dev/null +++ b/configs/spvcnn/README.md @@ -0,0 +1,44 @@ +# Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution + +> [Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution ](https://arxiv.org/abs/2007.16100) + + + +## Abstract + +Self-driving cars need to understand 3D scenes efficiently and accurately in order to drive safely. Given the limited hardware resources, existing 3D perception models are not able to recognize small instances (e.g., pedestrians, cyclists) very well due to the low-resolution voxelization and aggressive downsampling. To this end, we propose Sparse Point-Voxel Convolution (SPVConv), a lightweight 3D module that equips the vanilla Sparse Convolution with the high-resolution point-based branch. With negligible overhead, this point-based branch is able to preserve the fine details even from large outdoor scenes. To explore the spectrum of efficient 3D models, we first define a flexible architecture design space based on SPVConv, and we then present 3D Neural Architecture Search (3D-NAS) to search the optimal network architecture over this diverse design space efficiently and effectively. Experimental results validate that the resulting SPVNAS model is fast and accurate: it outperforms the state-of-the-art MinkowskiNet by 3.3%, ranking 1st on the competitive SemanticKITTI leaderboard. It also achieves 8x computation reduction and 3x measured speedup over MinkowskiNet with higher accuracy. Finally, we transfer our method to 3D object detection, and it achieves consistent improvements over the one-stage detection baseline on KITTI. + +
+ +
+ +## Introduction + +We implement SPVCNN with TorchSparse backend and provide the result and checkpoints on SemanticKITTI datasets. + +## Results and models + +### SemanticKITTI + +| Method | Lr schd | Mem (GB) | mIoU | Download | +| :--------: | :-----: | :------: | :--: | :----------------------: | +| SPVCNN-W16 | 15e | 3.9 | 61.9 | [model](<>) \| [log](<>) | +| SPVCNN-W20 | 15e | 4.2 | 62.7 | [model](<>) \| [log](<>) | +| SPVCNN-W32 | 15e | 5.4 | 63.4 | [model](<>) \| [log](<>) | + +**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 relatively dependent on random seeds, and if random seeds are not specified the model performance will be different (± 1.5 mIoU). + +## Citation + +```latex +@inproceedings{tang2020searching, + title={Searching efficient 3d architectures with sparse point-voxel convolution}, + author={Tang, Haotian and Liu, Zhijian and Zhao, Shengyu and Lin, Yujun and Lin, Ji and Wang, Hanrui and Han, Song}, + booktitle={Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XXVIII}, + pages={685--702}, + year={2020}, + organization={Springer} +} +``` diff --git a/configs/spvcnn/metafile.yml b/configs/spvcnn/metafile.yml new file mode 100644 index 0000000000..cc4b1444ef --- /dev/null +++ b/configs/spvcnn/metafile.yml @@ -0,0 +1,57 @@ +Collections: + - Name: SPVCNN + Metadata: + Training Techniques: + - AdamW + Architecture: + - SPVCNN + Paper: + URL: https://arxiv.org/abs/2007.16100 + Title: 'Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution' + README: configs/spvcnn/README.md + Code: + URL: + Version: v1.1.0rc3 + +Models: + - Name: spvcnn_w16_8xb2-15e_semantickitti + In Collection: SPVCNN + Config: configs/spvcnn/spvcnn_w16_8xb2-15e_semantickitti.py + Metadata: + Training Data: SemanticKITTI + Training Memory (GB): 3.9 + Training Resources: 8x A100 GPUs + Results: + - Task: 3D Semantic Segmentation + Dataset: SemanticKITTI + Metrics: + mIOU: 61.7 + Weights: + + - Name: spvcnn_w20_8xb2-15e_semantickitti + In Collection: SPVCNN + Config: configs/spvcnn/spvcnn_w20_8xb2-15e_semantickitti.py + Metadata: + Training Data: SemanticKITTI + Training Memory (GB): 4.2 + Training Resources: 8x A100 GPUs + Results: + - Task: 3D Semantic Segmentation + Dataset: SemanticKITTI + Metrics: + mIOU: 62.9 + Weights: + + - Name: spvcnn_w32_8xb2-15e_semantickitti + In Collection: SPVCNN + Config: configs/spvcnn/spvcnn_w32_8xb2-15e_semantickitti.py + Metadata: + Training Data: SemanticKITTI + Training Memory (GB): 5.4 + Training Resources: 8x A100 GPUs + Results: + - Task: 3D Semantic Segmentation + Dataset: SemanticKITTI + Metrics: + mIOU: 63.4 + Weights: diff --git a/model-index.yml b/model-index.yml index abc5ac939c..15653d2049 100644 --- a/model-index.yml +++ b/model-index.yml @@ -23,5 +23,6 @@ Import: - configs/smoke/metafile.yml - configs/ssn/metafile.yml - configs/votenet/metafile.yml + - configs/spvcnn/metafile.yml - configs/pv_rcnn/metafile.yml - configs/fcaf3d/metafile.yml