SE-SSD: Self-Ensembling Single-Stage Object Detector From Point Cloud (CVPR 2021) [Paper]
An accurate and fast single-stage 3D object detection framework on KITTI dataset.
Authors: Wu Zheng, Weiliang Tang, Li Jiang, Chi-Wing Fu.
original README.md
Ubuntu16.04 + CUDA10.2 + PyTorch1.6
注意：SE-SSD用到的apex库依赖pytorch1.6以上版本。
- pytorch版本问题
AT_CHECK相关的error， SE-SSD/det3d/core/iou3d应该是基于较老的pytorch版本，pytorch1.6不支持，需要将所有的AT_CHECK换成TORCH_CHECK。
- 训练报错
训练时报错TypeError: can't pickle _thread.RLock objects，修改SE-SSD/det3d/torchie/apis/train_sessd.py：
from det3d.models import build_detector
# model_ema = copy.deepcopy(model) # 这里报错 TypeError: can't pickle _thread.RLock objects
model_ema = build_detector(cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
if distributed:
model_ema = apex.parallel.convert_syncbn_model(model_ema)
model_ema = DistributedDataParallel(
model_ema.cuda(cfg.local_rank),
device_ids=[cfg.local_rank],
output_device=cfg.local_rank,
# broadcast_buffers=False,
find_unused_parameters=True,
)
else:
model_ema = model_ema.cuda()
- dataset
SE-SSD/det3d/datasets/dataset_factory.py中，把nuscenes和lyft相关行注释掉：
from .kitti import KittiDataset
# from .nuscenes import NuScenesDataset
# from .lyft import LyftDataset
dataset_factory = {
"KITTI": KittiDataset,
# "NUSC": NuScenesDataset,
# "LYFT": LyftDataset,
}
- 确保numpy安装的版本低于1.18，否则会报错
TypeError: 'numpy.float64' object cannot be interpreted as an integer：
pip install numpy==1.17.5 -i https://pypi.tuna.tsinghua.edu.cn/simple
PS：多留意pip是否是当前conda环境下的pip，以及pip的版本。
# <1> 我系统里已经安装了cuda10.2，只需要重定向
$ cd /usr/local
$ sudo rm cuda
$ sudo ln -s cuda10.2 cuda
# <2> 搭建虚拟环境
$ conda create -n sessd python=3.6
$ conda activate sessd
$ conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ #一般库的源
$ conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/ #pytorch, torchvision的源
$ conda install pip # 保证环境隔离
$ pip install --upgrade pip # 如果开了代理要先关掉
$ conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 #注意这里不加“-c pytorch”
$ PYTHONPATH=:/home/lina/anaconda3/envs/sessd/lib/python3.6/site-packages #我的bashrc里添加了多个路径给PYTHONPATH，这里删除其他环境影响
# <3> 安装SE-SSD
# <3.1> 安装依赖库
$ pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
# <3.2> 安装spconv，实测SE-SSD仅支持spconv1.×版本
$ git clone https://github.com/traveller59/spconv.git --recursive -b v1.2.1
$ cd spconv
$ sudo apt-get install libboost-all-dev
$ cmake --version # 确保cmake>= 3.13.2
$ python setup.py bdist_wheel
$ cd dist/
$ pip install spconv-1.2.1-cp36-cp36m-linux_x86_64.whl
# <3.3> 安装iou3d，先修改源码，把SE-SSD/det3d/core/iou3d下所有的AT_CHECK全部替换成TORCH_CHECK
$ conda install pyyaml
$ cd /home/lina/venti/SE-SSD/det3d/core/iou3d # AT_CHECK error
$ python setup.py install
# <3.4> 安装ifp-sample
$ mkdir third_party
$ cd third_party
$ git clone https://github.com/jackd/ifp-sample.git
$ pip install Cython -i https://pypi.tuna.tsinghua.edu.cn/simple
$ pip install -e ifp-sample
# <3.4> 安装apex
$ git clone https://github.com/NVIDIA/apex.git #apex是NIVIDIA用于混合精度训练的库，需要手动安装
$ cd apex
$ pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
# <3.4> 安装det3d
$ cd ../..
$ python setup.py build develop安装成功后的打印：
copying build/lib.linux-x86_64-3.6/det3d/ops/pointnet2/PN2.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/pointnet2
copying build/lib.linux-x86_64-3.6/det3d/ops/rroi_align/RotateRoIAlign.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/rroi_align
copying build/lib.linux-x86_64-3.6/det3d/ops/roipool3d/RoIPool3D.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/roipool3d
copying build/lib.linux-x86_64-3.6/det3d/ops/iou3d/IoU3D.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/iou3d
copying build/lib.linux-x86_64-3.6/det3d/ops/nms/nms.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/nms
copying build/lib.linux-x86_64-3.6/det3d/ops/sigmoid_focal_loss/sigmoid_focal_loss_cuda.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/sigmoid_focal_loss
copying build/lib.linux-x86_64-3.6/det3d/ops/syncbn/syncbn_gpu.cpython-36m-x86_64-linux-gnu.so -> det3d/ops/syncbn
Creating /home/lina/anaconda3/envs/sessd1/lib/python3.6/site-packages/det3d.egg-link (link to .)
Adding det3d 1.0rc0+b68068d to easy-install.pth file
Installed /home/lina/venti/SE-SSD
Processing dependencies for det3d==1.0rc0+b68068d
Searching for lyft-dataset-sdk==0.0.8
Best match: lyft-dataset-sdk 0.0.8
Adding lyft-dataset-sdk 0.0.8 to easy-install.pth file
...
...
...
Using /home/lina/anaconda3/envs/sessd1/lib/python3.6/site-packages
Searching for pycparser==2.21
Best match: pycparser 2.21
Adding pycparser 2.21 to easy-install.pth file
Using /home/lina/anaconda3/envs/sessd1/lib/python3.6/site-packages
Finished processing dependencies for det3d==1.0rc0+b68068dtrain阶段需要安装的库：
$ pip install ipdb -i https://pypi.tuna.tsinghua.edu.cn/simple$ cd SE-SSD/tools/
$ python create_data.py
$ python test.py --checkpoint se-ssd-model.pth
Evaluation official_AP_11: car AP(Average Precision)@0.70, 0.70, 0.70:
bbox AP:98.72, 90.10, 89.57
bev AP:90.61, 88.76, 88.18
3d AP:90.21, 86.25, 79.22
aos AP:98.67, 89.86, 89.16
car AP(Average Precision)@0.70, 0.50, 0.50:
bbox AP:98.72, 90.10, 89.57
bev AP:98.76, 90.19, 89.77
3d AP:98.73, 90.16, 89.72
aos AP:98.67, 89.86, 89.16
Evaluation official_AP_40: car AP(Average Precision)@0.70, 0.70, 0.70:
bbox AP:99.57, 95.58, 93.16
bev AP:96.70, 92.15, 89.74
3d AP:93.75, 86.18, 83.50
aos AP:99.52, 95.28, 92.69
car AP(Average Precision)@0.70, 0.50, 0.50:
bbox AP:99.57, 95.58, 93.16
bev AP:99.60, 95.92, 93.42
3d AP:99.59, 95.86, 93.36
aos AP:99.52, 95.28, 92.69$ cd SE-SSD/tools
$ python train.py # Single GPUUnfortunately, I cannot get a good trained model...
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent