Skip to content

Latest commit

 

History

History
121 lines (81 loc) · 4.61 KB

README.md

File metadata and controls

121 lines (81 loc) · 4.61 KB
Raw

pytorch-mask-rcnn

A pytorch implementation of Mask RCNN detection framework based on

This project supports single-GPU training of ResNet101-based Mask R-CNN (without FPN support). The purpose is to support the experiments in MAttNet, whose REFER dataset is a subset of COCO training portion. Thus our pre-trained model takes COCO_2014_train_minus_refer_valtest + COCO_2014_valminusminival images for training.

Prerequisites

  • Python 2.7
  • Pytorch 0.2 or higher
  • CUDA 8.0 or higher
  • requirements.txt

Preparation

  1. First of all, clone the code with refer API:
git clone --recursive https://github.com/lichengunc/mask-faster-rcnn
  1. Prepare data:
  • COCO: We use coco to name COCO's API as inheritance. Download the annotations and images into data/coco. Note the valminusminival and minival can be downloaded here.
git clone https://github.com/cocodataset/cocoapi data/coco
  • REFER: Follow the instructions in REFER to prepare the annotations for RefCOCO, RefCOCO+ and RefCOCOg.
git clone https://github.com/lichengunc/refer data/refer

  • ImageNet Weights: Find the resnet101-caffe download link from this repository, and download it as data/imagenet_weights/res101.pth.

  • coco_minus_refer: Make the coco_minus_refer annotation, which is to be saved as data/coco/annotations/instances_train_minus_refer_valtest2014.json

python tools/make_coco_minus_refer_instances.py

Compilation

As pointed out by ruotianluo/pytorch-faster-rcnn, choose the right -arch to compile the cuda code:

GPU model Architecture
TitanX (Maxwell/Pascal) sm_52
GTX 960M sm_50
GTX 1080 (Ti) sm_61
Grid K520 (AWS g2.2xlarge) sm_30
Tesla K80 (AWS p2.xlarge) sm_37

Compile the CUDA-based nms and roi_pooling using following simple commands:

cd lib
make

Training

Run by (notime as extra/tag)

./experiments/scripts/train_mask_rcnn_notime.sh 0 refcoco res101 notime
  • Train on COCO 2014 trainval35k minus refer_valtest, and test on minival (800k/1250k), 35.8 on detection and 30.7 on segmentation (nms).

Checking the training process by calling tensorboard, and check it at server.cs.unc.edu:port_number

tensorboard --logdir tensorboard/res101 --port=port_number

Evaluation

Run by (notime as extra/tag)

./experiments/scripts/test_mask_rcnn_notime.sh 0 refcoco res101 notime

Detection Comparison:

  • Compared with Faster R-CNN trained/evaluated on the same images.
Detection AP AP50 AP75
Faster R-CNN 34.1 53.7 36.8
Our Mask R-CNN 35.8 55.3 38.6

Segmentation Comparison:

  • We compare with Mask R-CNN implementation. Note this comparison is slightly unfair to ours, due to
  • We have fewer (~6,500) training images.
  • Our training is single GPU.
  • The shorter border length in our model is 600px instead of 800px.
Segmentation AP AP50 AP75
Original Mask R-CNN 32.7 54.2 34.0
Our Mask R-CNN 30.7 52.3 32.4

Pretrained Model

We provide the model we used in MAttNet for mask comprehension.

Download and put the downloaded .pth and .pkl files into output/res101/coco_2014_train_minus_refer_valtest+coco_2014_valminusminival/notime

Demo

  • Follow the ipython notebook of test/check_model.ipynb to test our model.