Skip to main content

BLIP: Bootstrapping Language-Image Pretraining for Unified Vision-Language Understanding

BLIP is a new vision-language model proposed by Microsoft Research Asia in 2022. It introduces a bootstrapping method to learn from noisy image-text pairs scraped from the web.

The BLIP Framework

BLIP consists of three key components:

  • MED - A multimodal encoder-decoder model that can encode images, text, and generate image-grounded text.
  • Captioner - Fine-tuned on COCO to generate captions for web images.
  • Filter - Fine-tuned on COCO to filter noisy image-text pairs.


The pretraining process follows these steps:
  1. Collect noisy image-text pairs from the web.
  2. Pretrain MED on this data.
  3. Finetune captioner and filter on the COCO dataset.
  4. Use captioner to generate new captions for web images.
  5. Filter noisy pairs using the filter model.
  6. Repeat the process by pretraining on a cleaned dataset.

This bootstrapping allows BLIP to learn from web-scale noisy data in a self-supervised manner.


Innovations in BLIP

Some interesting aspects of BLIP:

  • Combines encoder-decoder capability in one unified model (MED). Can encode, decode, and match images and text.
  • Careful pretraining objectives - image-text contrastive loss, matching loss, conditional language modeling loss.
  • Bootstraping loop to denoise image-text data.
  • Better performance than CLIP and ALIGN on vision-language tasks.

Results

BLIP achieves strong performance on vision-language benchmarks like text-image retrieval, visual question answering, etc. Some examples:

  • 83.5% accuracy on NLVR2 compared to 69.9% for CLIP.
  • 91.5% accuracy on Flickr30K image-text retrieval compared to 89.0% for ALIGN.
  • State-of-the-art on 8/14 vision-language tasks.

Limitations and Future Work

  • Bootstrapping can compound errors if the captioner or filter makes mistakes. Need careful fine-tuning.
  • Requires large supervised datasets (COCO) which can be expensive.
  • Can add iterative bootstrapping rounds to progressively improve the model.
  • Explore other modalities like video, audio, etc.

BLIP provides a scalable approach to learning joint vision-language representations from web data. The bootstrapping framework can pave the way for large multimodal models.


Comments

Post a Comment