search

ijkguo / mx-rcnn

Parallel Faster R-CNN implementation with MXNet.
Other
669 stars 292 forks source link

trained model with ResNet-101 #53

Closed wzkg2012 closed 7 years ago

wzkg2012 commented 7 years ago

I trained the model with ResNet-101 on VOC07+12 , but the performance is lower than what you reported, can you provide the trained model with resnet-101?

ijkguo commented 7 years ago

How much lower? I could certainly provide this model but I am working on something else so no time to verify.

wzkg2012 commented 7 years ago

I got 77.3% with ResNet-101 on VOC07+12 after 10 epoch, about 2% lower than you reported. Do more number of epochs get better results?

ijkguo commented 7 years ago

Our result is also from 10th epoch model.

This implementation follows the original paper closely, except the training counter (mxnet uses epochs; caffe uses iterations). We don't intend to be better but we have close results.