mitmul / chainer-fast-rcnn

Chainer folk of Fast R-CNN (Object Detection Method)

chainer-fast-rcnn

It aims at reproducing results of fast-rcnn using Chainer. It can be run only with GPU because roi_pooling_2d layer has only GPU implementation.

Requirements

• OpenCV 3.0 with python bindings
• Chainer 1.4
• fast-rcnn
• dlib v18.18

Create symlink

Create a symlink from the location of original fast-rcnn dir to this project's root dir. (The below line assumes a environment variable $FRCN_ROOT has a path to the fast-rcnn source dir.)

$ ln -s $FRCN_ROOT ./

Make sure that all steps written in the Installation (sufficient for the demo) section of README.md in fast-rcnn have been performed.

Convert model

Convert caffemodel to chainermodel.

$ python scripts/load_net.py

Test

First you should prepare a sample image, and then

$ python scripts/forward.py --img_fn sample.jpg --out_fn result.jpg

Result

'Overstekend wild' St. Janskerkhof Den Bosch © FaceMePLS (https://www.flickr.com/photos/faceme/5891724192)

chainer-fast-rcnn's Issues

problem with VGG.chainermodel

i have this problem when i use python scripts/load_net.py

Traceback (most recent call last):
File "scripts/load_net.py", line 25, in
setattr(vgg, name, layer)
File "/usr/local/lib/python2.7/dist-packages/chainer/function_set.py", line 38, in setattr
' %s' % value.name)
ValueError: given link is already registered to another chain by name conv1_1
how can i solve this ?

Can I train the network with this implementation?

Is there a way to train the fast rcnn network with this implementation?from the code I think currently is just loading the caffe network and evaluate the result. I'm looking forward to add faster rcnn support as well.

Also for roi layer looks like is not a problem to rewrite it with a forward_cpu and backward_cpu, the reason for gpu only is for performance considieration?