Large-scale Visual Relationship Understanding

Thie Repository is an implementation of the paper https://arxiv.org/pdf/1804.10660.pdf

Download the VRD Dataet

Download the data VRD dataset from here.

Pretrained Word Vectors

Download pretrained embeddings from here. Put it in the data/wordvectors folder. Folder structure is shown below.

Folder Structure

+ data 
    + VRD
        - json_dataset
        - sg_dataset
        
    + wordvectors  
        - GoogleNews-vectors-negative300.bin  
           

Train

./train.sh

Note

• All the weights will be saved to the snapshots folder
• To resume Training from any checkpoint, Use

--weight_path <path-to-model> 

Inference

python inference.py --input_image <path-to-image> 

python inference_batch.py --images_dir <path-to-dir>

To Do

Visual Genome Dataset, VGG BackBone, Resnet 152

Results

References

• https://github.com/facebookresearch/Large-Scale-VRD
• https://arxiv.org/pdf/1804.10660.pdf

License

This project is licensed under the MIT License