elainebao/py-faster-rcnn

Faster R-CNN (Python implementation) -- see https://github.com/ShaoqingRen/faster_rcnn for the official MATLAB version

Faster R-CNN with Imagenet dataset

This repository uses the python implementation of faster-rcnn py-faster-rcnn to train on imagenet dataset.

Difference

Most of the code is similar to the original py-faster-rcnn, expect the following scripts:

(We'll call the directory that you cloned Faster R-CNN into FRCN_ROOT)

1. $FRCN_ROOT/lib/datasets/imagenet.py & factory.py: add imagenet data interface.

The original py-faster-rcnn is trained on pascal_voc / coco dataset, there is no data interface for imagenet. So I add one for imagenet.

2. $FRCN_ROOT/models/imagenet: add models for imagenet.

Again the original py-faster-rcnn only provides prototxt for pascal_voc / coco dataset. The model I use for training imagenet is VGG16.

3. $FRCN_ROOT/experiments/cfgs/imagenet_faster_rcnn_end2end.yml: add imagenet config file. This should base on your path to imagenet dataset.

4. $FRCN_ROOT/experiments/scripts/faster_rcnn_end2end.sh: add imagenet train shell command.

5. $FRCN_ROOT/tools/demo_imagenet.py: add imagenet detection demo.

6. $FRCN_ROOT/lib/datasets/imagenet_eval.py: add imagenet evaluation.

To train and test with imagenet

1. first, download the training data and imagenet devkit

I use the training data from ILSVRC 2015, and arrange the data folder structure as follows:

2. Download pre-trained ImageNet models

Pre-trained ImageNet models can be downloaded for the three networks described in the paper: ZF and VGG16.

cd $FRCN_ROOT
./data/scripts/fetch_imagenet_models.sh

VGG16 comes from the Caffe Model Zoo, but is provided here for your convenience. ZF was trained at MSRA.

3. Modify config file

The file locates in $FRCN_ROOT/experiments/cfgs/imagenet_faster_rcnn_end2end.yml,
you can set path to imagenet dataset, etc. in this file.

4. Train faster-rcnn with Imagenet dataset

There are two training algorithms provided by the faster-rcnn NIPS 2015 paper, one is alternating optimization algorithm, and the other is approximate joint training algorithm.

Here I use the approximate joint training algorithm, as it resulting in faster (~ 1.5x speedup) training times and similar detection accuracy.

To train and test a Faster R-CNN detector using the approximate joint training method, use experiments/scripts/faster_rcnn_end2end.sh. Output is written underneath $FRCN_ROOT/output.

cd $FRCN_ROOT
./experiments/scripts/faster_rcnn_end2end.sh [GPU_ID] [NET] [DATASET] [--set ...]
# GPU_ID is the GPU you want to train on
# NET in {ZF, VGG_CNN_M_1024, VGG16} is the network arch to use
# DATASET is imagenet, or you can also use pascal voc, coco dataset.
# --set ... allows you to specify fast_rcnn.config options, e.g.

To detect with the model

1. Modify the model location in tools/demo_imagenet.py, modify the location of test prototxt and caffemodel.

python tools/demo_imagenet.py

For other issues, please visit the original py-faster-rcnn.