search

zhangliliang / RPN_BF

RPN+BF
Other
203 stars 75 forks source link

Caltech demo builds and runs but doesn't detect pedestrians #15

Closed JimKlingshirn closed 8 years ago

JimKlingshirn commented 8 years ago

Hi,

Thanks for sharing your code.

I've downloaded VGG16_caltech_final.zip, and built the code. When I run the demo, there are no pedestrians detected. The scores produced by the adaboost classifier are all very low (essentially zero). The region proposal values and scores seem reasonable. When I examine the "feat" vectors, most of the values are exactly zero. I'm guessing this is the problem. Any suggestions for debugging this would be very welcome.

My setup: Ubuntu 14.04, Cuda 8.0, CuDNN version 2, Nvidia GTX 1080, Matlab R2016b

Many thanks

zhangliliang commented 8 years ago

Hi,

The md5sum for the VGG16_caltech_final.zip should be equaled to df9e1aed34784a6be5f57ceb4a82ed03 .

Is it correct for you?

If so, I consider the main problem might be the different environment we used.

It might be solved after compling this mex file for matlab via this caffe version(https://github.com/zhangliliang/caffe/tree/RPN_BF) in your computer. Then replace that mex in the external/caffe/matlab/caffe_faster_rcnn/

JimKlingshirn commented 8 years ago

That md5sum is correct for me.

I did compile the mex file. In fact I never even tried the mex file you supplied. I'll try that.

I also suspect that I have a compatibility problem with my GPU/matlab/cuda combination. Apparently Matlab R2016b doesn't support Cuda 8.0. I'm going to try again with Cuda 7.5.

Thanks!

JimKlingshirn commented 8 years ago

It's working now. Instead of a new GPU (GTX 1080) with Cuda 8.0, I'm using an older Titan X with Cuda 7.5 and CuDNN version 2.

Thanks