hzr555/robotmaster-fast-detect

Ultra fast detection models for robomaster

Installation

Dowload, Install OpenVINO 2020.2 and finish all setup & GPU setup in document. Require Intel GPU or Intel Neural Compute Stick I don't recommend putting source /opt/intel/openvino/bin/setupvars.sh in .bashrc

Video demo

Model demo on official dataset

Model demo in the wild

Note: FPS reported in video is this model + our tracking pipeline's overall accuracy, which distributes compute between Neural Computing Stick and CPU. It's a couple times of the neural network itself, since you don't need to call inference every frame.

Model Optimization

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo.py --input_model CenterNet/onnx/rm_centernet_r18d4c6.onnx --data_type FP16 --batch 1 --mean_values [123.675,116.28,103.53] --scale_values [58.395,57.12,57.375]

python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo.py --input_model CenterNet/onnx/rm_centernet_r18d2c6.onnx --data_type FP16 --batch 1 --mean_values [123.675,116.28,103.53] --scale_values [58.395,57.12,57.375]

OpenVINO Inference (Intel NCS recommended)

source /opt/intel/openvino/bin/setupvars.sh export PYTHONPATH=$PYTHONPATH:. python3 Inference/openvino_inference

Paper

https://arxiv.org/pdf/1904.07850.pdf https://arxiv.org/abs/2012.07177

About data

You can find how we generated data in GenData. We only collected ~360 robot images from different angles.