Faster R-CNN

This directory contains code to evaluate the Faster R-CNN object detector described in the paper:

Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,
Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun,
Advances in Neural Information Processing Systems (NIPS), 2015

This code is based on the py-caffe implementation made available by Ross Girshick and the MatConvNet Fast R-CNN implementation by Hakan Bilen.

The pre-trained models released with the caffe code which have been imported into matconvnet and can be downloaded here, together with models trained directly with this code. Alternatively, you can train your own detector.

NOTE: The training code is still in the verfication process.

Demo

Running the faster_rcnn_demo.m script will download a model trained on pascal voc 2007 data and run it on a sample image to produce the figure below:

Functionality

There are scripts to evaluate models on the pascal voc and ms coco datasets (the scores produced by the pretrained models are listed on the model page). Training code is also provided to reproudce the pascal voc experiments described in the paper.

Dependencies

To simply run a detector in test mode, there are no additional dependencies. If you wish to train a new detector or improve detection efficiency, the following modules are required (these can be installed with vl_contrib):

• autonn - a wrapper module for matconvnet
• GPU NMS - a CUDA-based implementation of non-maximum supression

The effect of the CUDA NMS module is discussed below.

Performance

A comparison of the mean AP of the trained detectors is given here.

The Faster R-CNN pipeline makes heavy use of non-maximum suppression during training and inference. As a result, the runtime of the detector is significantly affected by the efficiency of the NMS function. A GPU version of non-maximum suppression can be found here, which can be compiled and added to your MATLAB path. Approximate benchmarks (they do not currently include the decoding of the raw predictions) of the code are given below on a Tesla M40 with a single item batch size:

mode	NMS (CPU)	NMS (GPU)
training	1.1 Hz	3.1 Hz
inference	6.7 Hz	8.3 Hz

Running the detector with on multiple GPUs produces a significant speed boost during inference, but currently only a minimal improvement during training (this may be addressed in future).

Multi-GPU code performance:

mode	Single GPU	2 GPUs
training	3.1 Hz	3.7 Hz
inference	7.5 Hz	15 Hz