search

zhengthomastang / 2018AICity_TeamUW

The winning method in Track 1 and Track 3 at the 2nd AI City Challenge Workshop in CVPR 2018 - Official Implementation
http://openaccess.thecvf.com/content_cvpr_2018_workshops/w3/html/Tang_Single-Camera_and_Inter-Camera_CVPR_2018_paper.html
550 stars 198 forks source link

Track1/6_SPD_EST issues #21

Open JoyceLiang-sudo opened 4 years ago

JoyceLiang-sudo commented 4 years ago

Thank you very much for sharing. I want to ask if speed detection is implemented in Python?

zhengthomastang commented 4 years ago

Sorry. Most of our implementations are in C++ and Matlab. But for our single-camera tracking part, there is a new version using PyTorch. You can also choose to use Python-based object detectors.

JoyceLiang-sudo commented 4 years ago

Sorry. Most of our implementations are in C++ and Matlab. But for our single-camera tracking part, there is a new version using PyTorch. You can also choose to use Python-based object detectors.

For the speed estimation,would you like to tell me how to get the camera parameters? Is it based on the camera that shot this video?

zhengthomastang commented 4 years ago

@tabu666 Yes. The camera parameters are different for each camera. You can use our provided camera calibration tool to compute them: https://github.com/zhengthomastang/2018AICity_TeamUW/tree/master/Track1/2_CAM_CAL

JoyceLiang-sudo commented 4 years ago

@zhengthomastang OK,thanks for your help!

bitasd commented 3 years ago

@zhengthomastang Hi Thomas! Could you please let me know: 1.why the speed has been scaled by 1.25, 1.05, 0.8, ets depending on the video (line 151 in main.cpp), and how these numbers have been drawn?

  1. why the accumulated distance between frames has been multiplied by 2.23694 (line 348 in main.cpp)
  2. why the number of used frames for the computation of average speed is set to 15 for some videos and 31 in some others Your response would help me out a lot! Thanks!!!
zhengthomastang commented 3 years ago

@bitasd Thanks for the questions. For (1) and (2), the scales on the speed values have been fine-tuned based on the actual speed limits and the evaluated performance on the test set. For (3), it is because the frame rates for the videos are different. We also considered the viewing perspectives to determine what a reliable range of frames for speed estimation should be.

zhengthomastang commented 3 years ago

@bitasd The constant 2.23694 (line 348 in main.cpp) is for converting meters/second to miles/hour. Only the parameters in (1) were tuned.

bitasd commented 3 years ago

You can also choose to use Python-based object detectors. , there is a new version using PyTorch But for our single-camera tracking part, there is a new version using PyTorch

Could you please point us to this python implementation? greatly appreciate it.

zhengthomastang commented 3 years ago

Hi. You can check out Faster R-CNN, SSD, YOLO, etc.