Semantic Object Accuracy for Generative Text-to-Image Synthesis

Code for our paper Semantic Object Accuracy for Generative Text-to-Image Synthesis (Arxiv Version) published in TPAMI 2020.

Summary in our blog post.

Semantic Object Accuracy (SOA) is a score we introduce to evaluate the quality of generative text-to-image models. For this, we provide captions from the MS-COCO data set from which the evaluated model should generate images. We then use a pre-trained object detector to check whether the generated images contain the object that was specified in the caption. E.g. when an image is generated from the caption a car is driving down the street we check if the generated image actually contains a car. For more details check section 4 of our paper.

We also perform a user study in which humans rate the images generated by several state-of-the art models trained on the MS-COCO dataset. We then compare the ranking obtained through our user study with the rankings obtained by different quantitative evaluation metrics. We show that popular metrics, such as e.g. the Inception Score, do not correlate with how humans rate the generated images, whereas SOA strongly correlates with human judgement.

Contents:

• Calculate SOA Scores
• Use Our Model (OP-GAN)

Calculate SOA Scores (Semantic Object Accuracy)

How to calculate the SOA scores for a model:

1. Go to SOA. The captions are in SOA/captions

   1. each file is named label_XX_XX.pkl describing for which labels the captions in the file are
   2. load the file with pickle

      •    import pickle 
           with open(label_XX_XX.pkl, "rb") as f:
               captions = pickle.load(f)
        

   3. each file is a list and each entry in the list is a dictionary containing information about the caption:

      •   [{'image_id': XX, 'id': XX, 'idx': [XX, XX], 'caption': u'XX'}, ...]
        

      • where 'idx': [XX, XX] gives the indices for the validation captions in the commonly used captions file from AttnGAN

2. Use your model to generate images from the specified captions

   1. each caption file contains the relevant captions for the given label
   2. create a new and empty folder
   3. use each caption file to generate images for each caption and save the images in a folder within the previously created empty folder, i.e. for each of the labels (0-79) there should be a new folder in the previously created folder and the folder structure should look like this
      • images
        • label_00 -> folder contains images generated from captions for label 0
        • label_01 -> folder contains images generated from captions for label 1
        • ...
        • label_79 -> folder contains images generated from captions for label 79
   4. each new folder (that contains generated images) should contain the string "label_XX" somewhere in its name (make sure that integers are formated to two digits, e.g. "0", "02", ...) -> ideally give the folders the same name as the label files
   5. generate three images for each caption in each file
      • exception: for label "00" (person) randomly sample 30,000 captions and generate one image each for a total of 30,000 images
   6. in the end you should have 80 folders in the folder created in the step (2.ii), each folder should have the string "label_XX" in it for identification, and each folder should contain the generated images for this label

3. Once you have generated images for each label you can calculate the SOA scores:

   1. Install requirements from SOA/requirements.txt (we use Python 3.5.2)
   2. download the YOLOv3 weights file and save it as SOA/yolov3.weights
   3. run python calculate_soa.py --images path/to/folder/created-in-step-2ii --output path/to/folder/where-results-are-saved --gpu 0

4. If you also want to calculate IoU values check the detailed instructions here

5. Calculating the SOA scores takes about 30-45 minutes (tested with a NVIDIA GTX 1080TI) depending on your hardware (not including the time it takes to generate the images)

6. More detailed information (if needed) here

Use Our Model (OP-GAN)

Dependencies

• python 3.8.5
• pytorch 1.7.1

Go to OP-GAN. Please add the project folder to PYTHONPATH and install the required dependencies:

conda env create -f environment.yml

Data

• MS-COCO:
  • download our preprocessed data (bounding boxes, bounding box labels, preprocessed captions), save to data/ and extract
    • the preprocessed captions are obtained from and are the same as in the AttnGAN implementation
    • the generateod bounding boxes for evaluating at test time were generated with code from the Obj-GAN
  • obtain the train and validation images from the 2014 split here, extract and save them in data/train/ and data/test/
  • download the pre-trained DAMSM for COCO model from here, put it into models/ and extract

Training

• to start training run sh train.sh gpu-ids where you choose which gpus to train on
  • e.g. sh train.sh 0,1,2,3
• training parameters can be adapted via code/cfg/dataset_train.yml, if you train on more/fewer GPUs or have more VRAM adjust the batch sizes as needed
• make sure the DATA_DIR in the respective code/cfg/cfg_file_train.yml points to the correct path
• results are stored in output/

Evaluating

• update the eval cfg file in code/cfg/dataset_eval.yml and adapt the path of NET_G to point to the model you want to use (default path is to the pretrained model linked below)
• run sh sample.sh gpu-ids to generate images using the specified model
  • e.g. sh sample.sh 0

Pretrained Models

• OP-GAN: download and save to models

Acknowledgement

• Code and preprocessed metadata for the experiments on MS-COCO are adapted from AttnGAN and AttnGAN+OP.
• Code to generate bounding boxes for evaluation at test time is from the Obj-GAN implementation.
• Code for using YOLOv3 is adapted from here, here, and here.

Citing

If you find our model useful in your research please consider citing:

@article{hinz2019semantic,
title     = {Semantic Object Accuracy for Generative Text-to-Image Synthesis},
author    = {Tobias Hinz and Stefan Heinrich and Stefan Wermter},
journal   = {arXiv preprint arXiv:1910.13321},
year      = {2019},
}