summary
- temporal convolutional networks (TCNs) を用いたPredictionの提案
Method
- 軽くて良さげな全体framework
- mobilenetなのがrealtimeを考慮していて良い感じ
- [[000361_uber_prediction_mtp]] MTP (Multiple Trajectory Prediction) をbaseに
- synthetic (合成) actor-centric rasterized image
- state input
- trackingの情報のuncertaintiesを使う
- temporal convolutional networks (TCNs)
- estimate position uncertainties
Experiment
- The model runs inference in
- 2.0 ms with batch size B = 1
- 7.4 ms with B = 32
- TensorRT acceleration (generation of images and uploading/downloading data to/from GPU excluded) with ONNX Runtime on an NVIDIA Geforce 1080 Ti.
- めっちゃ良さそう