search

GeekAlexis / FastMOT

High-performance multiple object tracking based on YOLO, Deep SORT, and KLT 🚀
MIT License
1.12k stars 256 forks source link

[Question] Performance and timing stats #223

Closed gokulas98 closed 2 years ago

gokulas98 commented 2 years ago

Hi, Thank you for your wonderful work. I have a few questions regarding performance. I am running it in Jetson Xavier AGX. These are the average fps I got while running outside container with default models MOT 17-13 : 29fps MOT 17-04 : 16fps MOT 17-03 : 10fps Are these expected values for AGX. or can it be improved?

  1. Does the performance in terms of FPS depends on the number of classes or only the object density?
  2. Do we have to implement ReID models for each classes. If so, would it affect the fps?
  3. Could you explain about the timing stats and what can we understand from it or what are common causes of it to increases? How are these calculated (is track time or any other calculated for each objects or frame)?
  4. Are all the timing stats calculated sequentially?
  5. Can you give a idea about what an ideal timing stats should be in the range of values?
  6. Will splitting the detection in to 2 models and running it parallelly increase the fps in jetson?
GeekAlexis commented 2 years ago

No, these are not normal. Please search past issues for runtime on Jetson.