ashual / scene_generation

A PyTorch implementation of the paper: Specifying Object Attributes and Relations in Interactive Scene Generation

Specifying Object Attributes and Relations in Interactive Scene Generation

A PyTorch implementation of the paper Specifying Object Attributes and Relations in Interactive Scene Generation

Paper

Specifying Object Attributes and Relations in Interactive Scene Generation
Oron Ashual¹, Lior Wolf^1,2
1 Tel-Aviv University, ² Facebook AI Research
The IEEE International Conference on Computer Vision (ICCV), 2019, (Oral)

Network Architechture

Youtube

Usage

1. Create a virtual environment (optional)

All code was developed and tested on Ubuntu 18.04 with Python 3.6 (Anaconda) and PyTorch 1.0.

conda create -n scene_generation python=3.7
conda activate scene_generation

2. Clone the repository

cd ~
git clone https://github.com/ashual/scene_generation.git
cd scene_generation

3. Install dependencies

conda install --file requirements.txt -c conda-forge -c pytorch

• install a PyTorch version which will fit your CUDA TOOLKIT

4. Install COCO API

Note: we didn't train our models with COCO panoptic dataset, the coco_panoptic.py code is for the sake of the community only.

cd ~
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI/
python setup.py install
cd ~/scene_generation

5. Train

$ python train.py

6. Encode the Appearance attributes

python scripts/encode_features --checkpoint TRAINED_MODEL_CHECKPOINT

7. Sample Images

python scripts/sample_images.py --checkpoint TRAINED_MODEL_CHECKPOINT --batch_size 32 --output_dir OUTPUT_DIR 

8. or Download trained models

Download these files into models/

9. Play with the GUI

The GUI was built as POC. Use it at your own risk:

python scripts/gui/simple-server.py --checkpoint YOUR_MODEL_CHECKPOINT --output_dir [DIR_NAME] --draw_scene_graphs 0

10. Results

Results were measured by sample images from the validation set and then running these 3 official scripts:

1. FID - https://github.com/bioinf-jku/TTUR (Tensorflow implementation)
2. Inception - https://github.com/openai/improved-gan/blob/master/inception_score/model.py (Tensorflow implementation)
3. Diversity - https://github.com/richzhang/PerceptualSimilarity (Pytorch implementation)
4. Accuracy - Training code is attached train_accuracy_net.py. A trained model is provided. Adding the argument --accuracy_model_path MODEL_PATH will output the accuracy of the objects.

Reproduce the comparison figure (Figure 3.)

Run this command

$ python scripts/sample_images.py --checkpoint TRAINED_MODEL_CHECKPOINT --output_dir OUTPUT_DIR

with these arguments:

• (c) - Ground truth layout: --use_gt_boxes 1 --use_gt_masks 1
• (d) - Ground truth location attributes: --use_gt_attr 1
• (e) - Ground truth appearance attributes: --use_gt_textures 1
• (f) - Scene Graph only - No extra attributes needed

Citation

If you find this code useful in your research then please cite

@InProceedings{Ashual_2019_ICCV,
    author = {Ashual, Oron and Wolf, Lior},
    title = {Specifying Object Attributes and Relations in Interactive Scene Generation},
    booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
    month = {October},
    year = {2019}
}

Acknowledgement

Our project borrows some source files from sg2im. We thank the authors.