bulid a simple faster rcnn model for VOC2007 base on Keras
This work aims to understand Faster RCNN architecture described by http://arxiv.org/pdf/1506.01497.pdf
Keras is a high-level deeplearning networks platform, it's easy to build a complex module such as Faster RCNN, which is highly helpful for architecture understanding.
Comparing with the orignal work, some charges are made for GPU-Mem-Limit or computing speed:
- a resize module is used instead of a roipooling layer
- fully connected layer reduce from 4096 to 2048
- a batchsize of 32 intead of 128 is feed after rpn layer
My softerware Env: Ubuntu 16.04 + tensorflow-gpu 1.4 + cuda 8.0 + CuDNN for cuda8.0
My hardware Env: GTX1060 6G
A jupyter-notebook based file is available for single step debug
some of the results:
This work is mainly build base on https://github.com/yhenon/keras-frcnn