Skip to content

Latest commit

 

History

History
91 lines (64 loc) · 2.83 KB

README.md

File metadata and controls

91 lines (64 loc) · 2.83 KB
Raw

Tiny Pascal VOC Instance Segmentation

To read the detailed , please, refer to CS_IOC5008_HW4
Based on https://github.com/matterport/Mask_RCNN

Hardware

The following specs were used to create the original solution.

  • Ubuntu 16.04 LTS
  • Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz
  • 1x NVIDIA TitanXp

Installation

All requirements should be detailed in requirements.txt. Using Anaconda is strongly recommended.

conda create -n mrnn python=3.6
source activate mrnn
pip install -r requirements.txt

Dataset Preparation

Following the Jupyter Notebook data loader in samples/VOC. And you have to create val.txt as validation images by your own.(Choose labels from train.txt)

Prepare Images

After execution of data loader, the data directory is structured as:

VOC
  +- VOCdevkit_Tiny
  |  +- VOC2012
  	|  +- Annotation
  	|  +- SegmentationClass
  	|  +- SegmentationObject
  	|  +- train_images
  	|  +- train.txt
  	|  +- val.txt
  +- data_loader.ipynb
  +- pascal_train.json
  +- voc.py

Training

You can set the config in voc.py (class VocConfig) The config template is in mrcnn/config.py

Model architecture(Callbacks, Optimizer, Regularizer)

You can set them in voc.py or model.py(recommend)

Train models

To train models, run following commands.

$ python voc.py train --dataset VOCdevkit_Tiny/ --model imagenet

The expected training times are:

Model GPUs Training Epochs Training Time
ResNet50 1x TitanXp 160 11 hours
ResNet101 1x TitanXp 160 14 hours

Pretrained models

We only can use ImageNet pretrained model for the fairness.

Training weights

The training weights are saved in logs directory.

Make Submission

Download the Test Image & json

Dataset GoogleDrive

At the directory samples/VOC/

$ wget https://drive.google.com/open?id=1VAhbaxLH7mjoPcJpoji5WijEoJy87wkk
$ wget https://drive.google.com/file/d/1s7OTV0rGW3AijyqSS6Tfl5IsDNGkiZWM/view?usp=sharing
$ tar xvf test_images.tar

Afterward, use demo.py to generate a json file.