codes for pedestrian detection
- Origin repo: https://github.com/polariseee/CSP-pedestrian-pytorch . Codes have been made some changes in this repo
- My best MR^-2 is 5.59%, a little better from the author's(5.84%)
- follow the origin repo to configure, and follow the UseIntro.txt to use my code.
refer to the Readme.md in file F2DNet
refer to the Readme.md in file Pedestron
- add a head branch(head center) with [conv 1×1×256] after "conv 3×3×256"
- add a head branch(head center + head scale + offset) with [conv 3×3×256 --> 3* conv 1×1×256] after feature map
- add a head branch(head center + head scale + offset) with [3* conv 1×1×256] after "conv 3×3×256"
- training and testing different methods using different loss weights
- best result is 5.42%MR^-2 with CSP_HeadCenter_3×3 and the loss weight of head center is 0.005. (a 0.17% improvement compared with the CSP code!)
Train.mp4
- Test: can choose images and modes to test,score and nms threshold can be changed,test results will be shown and printed
Test.mp4
- Track: can choose videos or open camera to track,should firstly detect a frame and then choose one tracking method(tradition methods) to track