Skip to content

Unofficial implementation of "Toward Realistic Image Compositing with Adversarial Learning (CVPR 2019)" in Jupyter Notebook

Notifications You must be signed in to change notification settings

SuhyeonHa/GCC-GANs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Toward Realistic Image Compositing with Adversarial Learning (CVPR 2019)

Unofficial implementation of "Toward Realistic Image Compositing with Adversarial Learning (CVPR 2019)" in PyTorch.

Example Results

Failure Cases

Networks

Dataset Preparation

First of all, download MS-COCO(train2014)

Data Pre-Processing

I followed the dataset generation process as described in the original paper.

Download validAnns_train.pkl, validAnns_dict_train.pkl, top5_dict_train.pkl, validAnns_val.pkl, validAnns_dict_val.pkl, and top5_dict_val.pkl (skipping procedure from 1 to 3).

When the phase is train,

1. "generate_valid_anns.ipynb": to filter out small objects and generate validAnns_train.pkl

2. "list-to-dict.ipynb": to generate validAnns_dict_train.pkl using validAnns_train.pkl

3. "top5-gcc-n0000.ipynb": to compare IoU between objects within the same category and pick top 5 items

  1. "mask-operation.ipynb": to save five different triplet images for a single object with validAnns_train.pkl and top5_dict_train.pkl

This procedures generate 678,685 training triplets(from 75,737 COCO objects). The total images hold 173G, so I won't upload them.

Training Dataset Examples

When the phase is test,

  1. "test-dataset-generation-top5-bg.ipynb": to save testing triplet images with validAnns_val.pkl and top5_dict_val.pkl

Testing images contain 363 triplets, and you can download them (52M) here.

Testing Dataset Examples

Getting Started

Installation

Clone this repo:

git clone https://github.com/SuhyeonHa/GCC-GANs
cd GCC-GANs

Before Training

Please make sure all directories are set right.

ann_dir = '/data/COCOdataset2017', # COCO dataset
data_dir = '/GCCdataset/alltypes', # GCC-GANs dataset
save_model_dir = '/GCC-GANs/models/', # Saving folder

About

Unofficial implementation of "Toward Realistic Image Compositing with Adversarial Learning (CVPR 2019)" in Jupyter Notebook

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published