NOTICE: This repo is no longer maintained. For an easy-to-use object detector that is actively maintained, I recommend considering the PyTorch Mask R-CNN implementation.
This repo contains code to download and evaluate the R-FCN object detector described in the paper:
"R-fcn: Object detection via region-based fully convolutional networks."
by Jifeng Dai, Li, Yi, Kaiming He, and Jian Sun (NIPS. 2016).
The implementation is based on the py-caffe code made
available by Yuwen Xiong. Pre-trained models released with the caffe code which have been imported into matconvnet can be downloaded here.
Running the rfcn_demo.m script will download a model trained on pascal voc 2007+2012 data and run it on a sample image to produce the figure below:
There are scripts to evaluate models on the pascal voc and mscoco datasets. The training code is still in the verfication process.
This module can be installed with the MatConvNet vl_contrib package manger. Due to the significant similarity in model design, this code re-uses part of the mcnFasterRCNN implementation. The following modules are required (these can also be installed with vl_contrib):
- autonn - a wrapper module for matconvnet
- GPU NMS - a CUDA-based implementation of non-maximum supression
- mcnFasterRCNN - MatConvNet Faster R-CNN
The scores produced by the pretrained models are listed on the model page. Running the detector with on multiple GPUs produces a significant speed boost during inference. Timings are shown below for the model based on the ResNet 50 and ResNet 101 models, averaged over a portion of the pascal 2007 test set using a Tesla M40 GPU with a single image minibatch. These benchmarks should be considered extremely apprxoimate - the variance on each image is high (due to differing input sizes), and they do not include the final short round of NMS :
| model | Single GPU | 2 GPUs |
|---|---|---|
| ResNet-50 | 8.3 Hz | 12.9 Hz |
| ResNet-101 | 4.2 Hz | 7 Hz |