Cross-modal Retrieval using Transformer Encoder Reasoning Networks

This project reimplements the idea from "Transformer Reasoning Network for Image-Text Matching and Retrieval". To solve the task of cross-modal retrieval, representative features from both modal are extracted using distinctive pipeline and then projected into the same embedding space. Because the features are sequence of vectors, Transformer-based model can be utilised to work best. In this repo, my highlight contribution is:

• Reimplement TERN module, which exploits the effectiveness of using Transformer on bottom-up attention features and bert features.
• Take advantage of facebookresearch's FAISS for efficient similarity search and clustering of dense vectors.
• Experiment various metric learning loss objectives from KevinMusgrave's Pytorch Metric Learning

The figure below shows the overview of the architecture

Datasets

• I trained TERN on Flickr30k dataset which contains 31,000 images collected from Flickr, together with 5 reference sentences provided by human annotators for each image. For each sample, visual and text features are pre-extracted as numpy files

• Some samples from the dataset:

Images	Captions
	1. An elderly man is setting the table in front of an open door that leads outside to a garden. 2. The guy in the black sweater is looking onto the table below. 3. A man in a black jacket picking something up from a table. 4. An old man wearing a black jacket is looking on the table. 5. The gray-haired man is wearing a sweater.
	1. Two men are working on a bicycle on the side of the road. 2. Three men working on a bicycle on a cobblestone street. 3. Two men wearing shorts are working on a blue bike. 4. Three men inspecting a bicycle on a street. 5. Three men examining a bicycle.

Execution

• Installation

pip install -r requirements.txt
apt install libomp-dev
pip install faiss-gpu

• Specify dataset paths and configuration in the config file

• For training

PYTHONPATH=. python tools/train.py 

• For evaluation

PYTHONPATH=. python tools/eval.py \
                --top_k= <top k similarity> \
                --weight= <model checkpoint> \

• For inference
  • See tools/inference.py script

Notebooks

• Inference TERN on Flickr30k dataset
• Use FasterRCNN to extract Bottom Up embeddings
• Use BERT to extract text embeddings

Results

• Validation m on Flickr30k dataset (trained for 100 epochs):

Model	Weights	i2t/R@10	t2i/R@10
TERN	link	0.5174	0.7496

• Some visualization

Query text: Two dogs are running along the street

Query text: The woman is holding a violin

Query text: Young boys are playing baseball

Query text: A man is standing, looking at a lake

Paper References

@misc{messina2021transformer,
      title={Transformer Reasoning Network for Image-Text Matching and Retrieval}, 
      author={Nicola Messina and Fabrizio Falchi and Andrea Esuli and Giuseppe Amato},
      year={2021},
      eprint={2004.09144},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{anderson2018bottomup,
      title={Bottom-Up and Top-Down Attention for Image Captioning and Visual Question Answering}, 
      author={Peter Anderson and Xiaodong He and Chris Buehler and Damien Teney and Mark Johnson and Stephen Gould and Lei Zhang},
      year={2018},
      eprint={1707.07998},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@article{JDH17,
  title={Billion-scale similarity search with GPUs},
  author={Johnson, Jeff and Douze, Matthijs and J{\'e}gou, Herv{\'e}},
  journal={arXiv preprint arXiv:1702.08734},
  year={2017}
}

Code References

• https://github.com/facebookresearch/faiss
• https://github.com/KevinMusgrave/pytorch-metric-learning
• https://github.com/huggingface/transformers
• https://github.com/airsplay/py-bottom-up-attention