search

tensorturtle / classy-sort-yolov5

Ready-to-use realtime multi-object tracker that works for any object category. YOLOv5 + SORT implementation.
GNU General Public License v3.0
110 stars 29 forks source link

trying to use sort on the jetson target, any way to reduce the time of performance? #12

Closed YoungjaeDev closed 2 years ago

YoungjaeDev commented 2 years ago

Sort takes about 5-9 times more than the yolo5 detect I think it's due to CPU processing, is there any way to reduce it?

Thank you

tensorturtle commented 2 years ago

Hi @youngjae-avikus , I totally relate with your issue. I have tried to run it on Jetson. You are right, SORT is somewhat slow on the CPU. If you're willing to sacrifice precision, I recommend IOUTracker C++ implementations of SORT are probably faster, but I have not tested these.

YoungjaeDev commented 2 years ago

Is DeepSort the same because it is Feature info + Sort after all? I'm curious about how to attach the tracker at Real time. Thank you

tensorturtle commented 2 years ago

If you're looking for speed, DeepSORT won't help you - it runs a separate neural network on top of the object detection. Check out QSORT, which I think is close to a drop-in replacement for SORT.

YoungjaeDev commented 2 years ago

thank you! I wonder that sort is kf+iou tracker? Predicting with Kf and doing association with iou, this is how I think. is it right? I think iou tracker is more faster!

tensorturtle commented 2 years ago

@youngjae-avikus Yes your understanding is correct- I suggest you take a look at the code, it's relatively clearly written and commented.