NOVA-3D is the industry's first all-in-one optimization, acceleration and deployment tool that is specifically created for 3D point cloud algorithms. The self-developed solution by NOVAUTO features a complete point cloud operator library, a pruning and quantization tool, a reference model library and an inference acceleration library. NOVA-3D effectively solves the major pain point of 3D point cloud algorithms that are usually too complex for deployment on embedded systems while ensuring high performance and real-time inference of the models. The high level of automation of NOVA-3D greatly reduces project complexity and helps our customers to realize their AI solutions fast, at low costs and with unparalleled performance.
The model library contains numerous state-of-the-art models (voxel-based, point-based, view-based, LiDAR-image, etc.) that can all be deployed in a unified manner on NVIDIA's embedded platforms such as Jetson Xavier or Jetson Orin.
A great example for the effectiveness of NOVA-3D is the included state-of-the-art model CIA-SSD. To give you a first-hand impression, this repository provides everything necessary for the deployment of the NOVA-3D CIA-SSD model on your Jetson AGX Xavier platform (jetpack4.6) with evaluation on the KITTI dataset: We provide two optimized ONNX models (floating point and quantization), the required TensorRT inference tool and the KITTI evaluation code for seamless validation of the results.
With NOVA-3D, our customers can realize their cutting-edge LiDAR perception solutions with accelerated deployment on NVIDIA's embedded platforms such as Jetson Xavier or Jetson Orin. If you are interested, please contact us.
On your NVIDIA Jetson AGX Xavier(jetpack4.6), compile the program:
$ mkdir build
$ cd build
$ cmake ..
$ make -j8Set up your LD_LIBRARY_PATH:
$ export LD_LIBRARY_PATH=$(pwd)/nova3d_lib/:$LD_LIBRARY_PATHSet the mode in the config.xml file and run the program:
# fp32 or fp16: set MODE=0(fp32) or MODE=1(fp16)
$ ./unit_test --onnx_path ../model/ciassd_float.onnx --data_path ../validation_set/
# int8: set MODE=2(int8)
$ ./unit_test --onnx_path ../model/ciassd_quant.onnx --data_path ../validation_set/Evaluate the results (requires python3):
python3 eval_kitti.py --pre_path ../output/NOVA-3D CIA-SSD on NVIDIA Jetson AGX Xavier:
| 3D AP_11 (%) | 3D AP_40 (%) | inference time (ms) | |
|---|---|---|---|
| FP32 | 80.0 | 83.8 | 123.2 |
| FP16 | 80.0 | 83.9 | 91.2 |
| INT8 | 79.45 | 83.12 | 65.13 |
To reproduce our results, use the provided KITTI validation set. The point cloud data is reduced to the area visible by the camera.
Besides, we also get the latency on Jetson AGX ORIN:
| inference time (ms) | |
|---|---|
| FP32 | 40.9 |
| FP16 | 35.6 |
| INT8 | 32.1 |
Hao Liu, Zhongyuan Qiu, Yifei Chen, Yali Zhao
This project is licensed under the Apache license 2.0 License - see the LICENSE file for details.
The code is developed based on TensorRT. Thanks for previous works CIA-SSD.
If you have any question or suggestion about this repo, please contact us(zhongyuan.qiu@novauto.com.cn, hao.liu@novauto.com.cn, yifei.chen@novauto.com.cn).