search

xw-hu / SINet

IEEE Transactions on Intelligent Transportation Systems (TITS), 2019
Other
80 stars 22 forks source link

Input Resolution Influence on mAP #4

Closed skywalker9096 closed 5 years ago

skywalker9096 commented 6 years ago

hello @xw-hu , I followed you instructions and trained the model myself and get good performance(mAP on Car: 89.63), but when I use smaller input size on the same model, the mAP drops to 79.40/56.42. any idea what happens? I am rather confused.

model resolution Car Ped Cyc FPS
mscnn 1280x384 75.43 78.95 77.04 11
mscnn 2560x756 91.38 - - 3
SINet 1920x576 89.63 72.26 61.72 21
SINet 1280x384 79.40 65.37 44.47 37
SINet-Car 864x256 56.42 - - 59

I use 3712 images for training and 3769 images for validation.

xw-hu commented 6 years ago

I guess you forgot to adjust some parameters (such as the size of anchor, the parameters of ROISplit Layer). Since the original parameters are calculated for Car detection, that is why the accuracy of Ped and Cyc is not as good as Car. You may re-calculate these parameters, and I also provide the matlab files on parameter calculation for your reference.