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