Segmentation_model: Internimage Ops: DCNv3 Transformer: LSSviewTransformer + bevpoolv2_cpu
This repository will generate result.pkl fusion data
The InternImage backbone with dcnv3 has cpu build capabilities also bevpool and lssview transformer has cpu build capabilities This readme provides instructions for setting up and using the CPU version of the SFusion model with DCNV3, LSSViewTransformer, and DepthNet on the NuScenes dataset. SFusion is a sensor fusion model designed to work with radar and camera data for 3D Segmentation and depth estimation.
mmcv=1.4.0=pypi_0
mmcv-full=1.4.0=pypi_0
mmdet=2.28.1=dev_0
mmengine=0.8.2=pypi_0
mmsegmentation=0.30.0=pypi_0
onnx=1.14.1=pypi_0
onnxruntime=1.16.0=pypi_0
pytorch=1.9.0=py3.9_cpu_0To run this model on CPU, you will need Either
- A machine with a CPU that supports AVX2 instructions
- A machine with a CPU that has minimum 16GB RAM.
- Prepare nuScenes dataset(v1.0trainval) Download nuScenes 3D detection data and unzip all zip files. The folder structure should be organized as follows before our processing.
HVDetFusion-cpu_build
├── mmdet3d
├── tools
├── configs
├── data
│ ├── nuscenes
│ │ ├── maps
│ │ ├── samples
│ │ ├── sweeps
│ │ ├── v1.0-test
| | ├── v1.0-trainval
Setup and Inference(Before running the docker make sure the Dataset is prepared based on the above dir structure)
docker build
python HVDet_infer.py configs/hvdet/HVDetInfer_sim.py tools/convter2onnx/onnx_output --fuse-conv-bn --eval bbox # --offline_eval --out ./res_pkl/test.pkl
This work is built on the open-sourced HVDetFusion, BevDet,BevDepth and the published code of CenterFusion.
This project is released under the Apache 2.0 license.