Join 10350+ others. No spamming.
I promise!

Follow us at github.



facebookresearch / multipathnet


A Torch implementation of the object detection network from "A MultiPath Network for Object Detection" (


MultiPath Network training code

The code provides functionality to train Fast R-CNN and MultiPath Networks in Torch-7.
Corresponding paper: A MultiPath Network for Object Detection


If you use MultiPathNet in your research, please cite the relevant papers:

    author = {S. Zagoruyko and A. Lerer and T.-Y. Lin and P. O. Pinheiro and S. Gross and S. Chintala and P. Doll{\'{a}}r},
    title = {A MultiPath Network for Object Detection},
    booktitle = {BMVC}
    year = {2016}


  • Linux
  • NVIDIA GPU with compute capability 3.5+


The code depends on Torch-7, fb.python and several other easy-to-install torch packages.
To install Torch, follow
Then install additional packages:

luarocks install inn
luarocks install torchnet
luarocks install fbpython
luarocks install class

Evaluation relies on COCO API calls via python interface, because lua interface doesn't support it. Lua API is used to load annotation files in *json to COCO API data structures. This doesn't work for proposal files as they're too big, so we provide converted proposals for sharpmask and selective search in torch format.

First, clone

git clone

Then install LuaAPI:

cd coco
luarocks make LuaAPI/rocks/coco-scm-1.rockspec

And PythonAPI:

cd coco/PythonAPI

You might need to install Cython for this:

sudo apt-get install python-pip
sudo pip install Cython

You will have to add the path to PythonAPI to PYTHONPATH. Note that this won't work with anaconda as it ships with it's own libraries which conflict with torch.

EC2 installation script

Thanks to @DeegC there is scripts/ script for quick EC2 setup.

Data preparation

The root folder should have a folder data with the following subfolders:


models folder should contain AlexNet and VGG pretrained imagenet files downloaded from here. ResNets can resident in other places specified by resnet_path env variable.

annotations should contain *json files downloaded from There are *json annotation files for PASCAL VOC, MSCOCO, ImageNet and other datasets.

proposals should contain *t7 files downloaded from here We provide selective search VOC 2007 and VOC 2012 proposals converted from and SharpMask proposals for COCO 2015 converted from, which can be used to compute proposals for new images as well.

Here is an example structure:

|-- annotations
|   |-- instances_train2014.json
|   |-- instances_val2014.json
|   |-- pascal_test2007.json
|   |-- pascal_train2007.json
|   |-- pascal_train2012.json
|   |-- pascal_val2007.json
|   `-- pascal_val2012.json
|-- models
|   |-- caffenet_fast_rcnn_iter_40000.t7
|   |-- imagenet_pretrained_alexnet.t7
|   |-- imagenet_pretrained_vgg.t7
|   `-- vgg16_fast_rcnn_iter_40000.t7
`-- proposals
    |-- VOC2007
    |   `-- selective_search
    |       |-- test.t7
    |       |-- train.t7
    |       |-- trainval.t7
    |       `-- val.t7
    `-- coco
        `-- sharpmask
            |-- train.t7
            `-- val.t7

Download selective_search proposals for VOC2007:


Download sharpmask proposals for COCO:


As for the images themselves, provide paths to VOCDevkit and COCO in config.lua

Running DeepMask with MultiPathNet on provided image

We provide an example of how to extract DeepMask or SharpMask proposals from an image and run recognition MultiPathNet to classify them, then do non-maximum suppression and draw the found objects.

  1. Clone DeepMask project into the root directory:
git clone
  1. Download DeepMask or SharpMask network:
cd data/models
# download SharpMask based on ResNet-50
wget -O sharpmask.t7
  1. Download recognition network:
cd data/models
# download ResNet-18-based model trained on COCO with integral loss
  1. Make sure you have COCO validation .json files in data/annotations/instances_val2014.json

  2. Pick some image and run the script:

th demo.lua -img ./deepmask/data/testImage.jpg

And you should see this image:

iterm2 4jpuod lua_khbaaq

See file demo.lua for details.


The repository supports training Fast-RCNN and MultiPath networks with data and model multi-GPU paralellism. Supported base models are the following:


To train Fast-RCNN on VOC2007 trainval with VGG base model and selective search proposals do:

test_nsamples=1000 model=vgg ./scripts/

The resulting mAP is slightly (~2 mAP) higher than original Fast-RCNN number. We should mention that the code is not exactly the same as we improved ROIPooling by fixing a few bugs, see


To train MultiPathNet with VGG-16 base model on 4 GPUs run:

train_nGPU=4 test_nGPU=1 ./scripts/

Here is a graph visualization of the network (click to enlarge):


To train ResNet-18 on COCO do:

train_nGPU=4 test_nGPU=1 model=resnet resnet_path=./data/models/resnet/resnet-18.t7 ./scripts/



We provide original models from Fast-RCNN paper converted to torch format here:

To evaluate these models run:

model=data/models/caffenet_fast_rcnn_iter_40000.t7 ./scripts/
model=data/models/vgg_fast_rcnn_iter_40000.t7 ./scripts/


Evaluate fast ResNet-18-based network trained with integral loss on COCO val5k split (resnet18_integral_coco.t7 89MB):

test_nGPU=4 test_nsamples=5000 ./scripts/

It achieves 24.4 mAP using 400 SharpMask proposals per image:

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.244
Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.402
Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.268
Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.078
Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.266
Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.394
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.249
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.368
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.377
Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.135
Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.444
Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.561