Technical Name 通曉運算量之AI模型架構優化、即時運算實現與資料集標註系統
Project Operator National Chung Hsing University
Project Host 陳冠宏
1. CNN model architecture optimization techniques: Propose an Agilev3-L architecture, with a performance index of mAP@50:95FPS:FP16 up to 171.79
2. GOP-mode acceleration scheme for real time inference: employ GOP-modeTRACKING algorithms to enhance the processing rate. 
3. Propose a rapid data labeling system HiTag.
Technical Film
Scientific Breakthrough
1. CNN model architecture optimization techniques: Propose a new model architecture Agilev3-L achieving a composite performance index of 171.79 (yolov3 is 135.24, yolov4 is 163.83)
2. GOP based acceleration scheme for real time inference: Split the video input into I-framesP-frames. Only I-frames are predicted, while the P-frames results are obtained by tracking. For implementation of the Agilev3-L model on a Jetson Nano platform, the FPS is improved from the 3.91 to 27.09, indicating a 592 enhancement. AP@50 performance drops slightly from 85.21 to 84.69.
Industrial Applicability
Systematic CNN modelarchitecture optimization techniques are developed to facilitate real time inference at edge sides. For the proposed Agilev3-L model, the implementation on a Jetson Nano platform can reach a processing rate of  27.09 FPS. We have also developed an auto-labeling system to expedite the tedioustime consuming ground truth bounding box labeling.
Matching Needs
Keyword Neural Network Object Detection Image Recognition GFLOPS Object Tracking Image Preprocessing Group of Pictures Model Predict Tracking Datasets Automatic Labeling
  • Contact