Open guo-pu opened 10 months ago
When the number of objects increases, the regression reasoning time increases exponentially.
It's true that when object increases it will cost a loooong time, but when we export dynamic batch size model, we can infer objects once a time.
Hello, the model is in two stages when reasoning, first "detect" the 2d information of the target, then "regression" 3D information, and finally combine geometric constraints to generate 3D box information. In regression reasoning, each detected object is regressed separately, so there is a hidden danger: When the number of objects increases, the regression reasoning time increases exponentially.
I wonder if I can integrate all the detected objects into a single picture, so that the time of regression reasoning is fixed and the speed of reasoning is improved. Or we can improve the detection head of yolov5, add parallel branches (3D size, orientation, etc.) to the original detection 2d box position and category, and realize one-stage model reasoning.