Awesome

PolygonRNN++

This is the official PyTorch reimplementation of Polygon-RNN++ (CVPR 2018). This repository allows you to train new Polygon-RNN++ models, and run our demo tool on local machines. For technical details, please refer to:

Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++
David Acuna*, Huan Ling*, Amlan Kar*, Sanja Fidler (* denotes equal contribution)
CVPR 2018
[Paper] [Video] [Project Page] [Demo]
<img src = "Docs/model.png" width="56%"/> <img src = "Docs/polydemo.gif" width="42%"/>

Where is the code?

To get the code, please signup here. We will be using GitHub to keep track of issues with the code and to update on availability of newer versions (also available on website and through e-mail to signed up users).

If you use this code, please cite:

@inproceedings{AcunaCVPR18,
title={Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++},
author={David Acuna and Huan Ling and Amlan Kar and Sanja Fidler},
booktitle={CVPR},
year={2018}
}

@inproceedings{CastrejonCVPR17,
title = {Annotating Object Instances with a Polygon-RNN},
author = {Lluis Castrejon and Kaustav Kundu and Raquel Urtasun and Sanja Fidler},
booktitle = {CVPR},
year = {2017}
}

Contents

1. Reproduction Results
2. Environment Setup
3. Tool
   1. Backend
   2. Frontend
4. Testing Models
5. Training Models
   1. Data
   2. Training MLE Model
   3. Training RL Model
   4. Training Evaluator
   5. Training GGNN

Results

These are the reproduction results from this repository as compared to the paper

Note: Benchmarked forward pass speed for the tool (with 5 first points and 1 beam size) is 0.3 seconds per interaction on a TitanXp

Note: Shared Encoder refers to sharing the Resnet between the graph network and the convLSTM network. In the original paper, the two networks were kept separate.

Environment Setup

All the code has been run and tested on Ubuntu 16.04, Python 2.7.12, Pytorch 0.4.0, CUDA 9.0, TITAN X/Xp and GTX 1080Ti GPUs

• Get code after signing up
• Go into the downloaded code directory

cd <path_to_downloaded_directory>

• Setup python environment

virtualenv env
source env/bin/activate
pip install -r requirements.txt

• Add the project to PYTHONPATH

export PYTHONPATH=$PWD

Tool

• Setup your environment
• Download the MLE+RL+Evaluator+GGNN model after getting access by signing up

Backend

• Launch backend (flask server) with,

python Tool/tool.py --exp Experiments/tool.json --reload <path_to_model> --port <port> --image_dir Tool/frontend/static/img/

Frontend

• Edit Tool/frontend/static/js/polygon.js and change globalFolder to the appropriate directory based on where you cloned the repository.
• With python2.7, run

cd Tool/frontend/
python -m SimpleHTTPServer

• On your browser, navigate to localhost:8000. You should see a page like <img src = "Docs/tool.png" width="100%"/>

Note: Replace SimpleHTTPServer with http.server if you are using python3 for the server

Note: You can setup your own image directory by editing Tool/frontend/static/js/polygon.js and passing that path to Tool/tool.py from the command line. This image directory MUST contain the pre-defined images that are defined in Tool/frontend/index.html

Testing Models

• Setup your environment
• Download pretrained models after getting access by signing up

python Scripts/prediction/generate_annotation.py --exp <path_to_corresponding_experiment> --reload <path_to_checkpoint> --output_dir <path_to_store_predictions>

• If you are testing a RL checkpoint, then the experiment file would correspond be the RL experiment file, and similarly for other stages of the model
• You can check predicted/GT masks for every instance in the output_dir
• To get scores, run

python Scripts/get_scores.py --pred <path_to_preds> --output <path_to_file_to_save_results>

Training Models

Data

Cityscapes

• Download the Cityscapes dataset (leftImg8bit_trainvaltest.zip) from the official website [11 GB]
• Our processed annotation files are included in the download file you get after signing up
• From the root directory, run the following command with appropriate paths to get the annotation files ready for your machine

python Scripts/data/change_paths.py --city_dir <path_to_downloaded_leftImg8bit_folder> --json_dir <path_to_downloaded_annotation_file> --out_dir <output_dir>

Custom Dataset

To train on your custom datasets, you have one of two options:

• Prepare annotation data similar to our annotation files and use our default DataProvider
• Implement your own DataProvider following the cityscapes implementation for your own data

Training

• Setup your environment
• Download the pre-trained Pytorch Resnet-50 from here
• Note - While resuming training, always resume from end of epoch checkpoints to produce reproducible results!

Training MLE model

• Edit the experiment file at Experiments/mle.json and change paths for your machine
• From the root directory, run

python Scripts/train/train_ce.py --exp Experiments/mle.json --resume <optional_if_resuming_training>

• You can view progress on Tensorboard (logs are at <experiment_dir>/logs/)

Training RL model

• Edit the experiment file at Experiments/rl.json and change paths for your machine
• In the experiment file, set xe_initializer to the best MLE model
• From the root directory, run

python Scripts/train/train_rl.py --exp Experiments/rl.json --resume <optional_if_resuming_training>

• Note - You might have to play with hyperparameters a bit to achieve stable training, especially temperature, lr and lr_decay

Training Evaluator

• Edit the experiment file at Experiments/evaluator.json and change paths for your machine
• In the experiment file, set xe_initializer to the best RL model
• From the root directory, run

python Scripts/train/train_evaluator.py --exp Experiments/evaluator.json --resume <optional_if_resuming_training>

Training GGNN

• Edit the experiment file at Experiments/ggnn.json and change paths for your machine
• In the experiment file, set xe_initializer to the best Evaluator model
• From the root directory, run

python Scripts/train/train_ggnn.py --exp Experiments/ggnn.json --resume <optional_if_resuming_training>