Experimental Torch7 implementation of RCNN for Object Detection with a Region Proposal Network
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
config
models Add SpatialDropout, BatchNormalization, input centering and scaling Oct 19, 2015
.gitignore Initial commit Sep 17, 2015
Anchors.lua Reenable contrastive normalization Oct 19, 2015
BatchIterator.lua Add SpatialDropout, BatchNormalization, input centering and scaling Oct 19, 2015
Detector.lua Reenable contrastive normalization Oct 19, 2015
LICENSE Initial commit Sep 17, 2015
Localizer.lua Fix anchor index calculation Sep 21, 2015
README.md Update paper url Jun 13, 2016
Rect.lua Add basic augmented image loading (scale, crop, vflip, hflip) Sep 24, 2015
create-duplo-traindata.lua Add image net detection import Oct 7, 2015
create-imagenet-traindata.lua Add negative examples (background) images of ILSVRC2015 (imagenet) data Oct 7, 2015
main.lua Add SpatialDropout, BatchNormalization, input centering and scaling Oct 19, 2015
nms.lua Add final per class non-maximum supperession Sep 21, 2015
objective.lua Add missing config files Oct 16, 2015
utilities.lua Restructure model creation and configuration Oct 15, 2015

README.md

faster-rcnn

This is an experimental Torch7 implementation of Faster RCNN - a convnet for object detection with a region proposal network. For details about R-CNN please refer to the paper Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks by Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun.

Work in progress

Status: Basic detection in my personal environment works. A 'small' network is used that can be trained on a 4 GB GPU with 800x450 images. Began experimenting with ImageNet: create-imagenet-traindat.lua can be used to create a training data file for the ILSVRC2015 dataset.

Todo:

  • [!] regularly evaluate net during traning to compute test-set loss
  • generate training graph with gnuplot
  • add final per class non-maximum suppression to generate final proposals (already included but eval code rewrite still pending)
  • remove hard coded path, create full set of command line options
  • add parameters to separately enable/disable training of bounding box proposal-network and fine-tuning + classification.

Experiments to run:

  • test smaller networks
  • 6x6 vs. 7x7 classification ROI-pooling output size
  • impact of RGB, YUV, Lab color space
  • test relevance of local contrast normalization

References / Review / Useful Links