日韩精品久久久毛片一区二区,亚洲国产韩国一区二区,国产成人在线免费观看

Abstract

LLMs have demonstrated remarkable abilities at interacting with humans through language, especially with the usage of instruction-following data. Recent advancements in LLMs, such as MiniGPT-4, LLaVA, and X-LLM, further enlarge their abilities by incorporating multi-modal inputs, including image, video, and speech. Despite their effectiveness at generating precise and detailed language understanding of the given modality signal, these LLMs give up the ability to ground specific parts of inputs, thus only constructing a coarse-grained mapping. However, explicit and informative correspondence between text and other modalities will not only improve the user experience but also help to expand the application scenario of multi-modal LLMs.

BuboGPT Architecture . We build a multi-modal LLM, BuboGPT for multi-modal understanding including image, audio and text by learning a common semantic space and further explore the fine-grained relation between different visual objects and different modalities.
Multimodal Instruct Data. We construct a high-quality multi-modal instruction-tuning dataset including fine-grained audio descriptions and cross-modal sound localization, and introduce both positive and negative image-audio pairs for semantic matching to facilitate the cross-modal understanding..

BuboGPT Architecture

As the figure shown, we perform joint multi-modal understanding and chatting for text, vision and audio, which is achieved by learning a shared representation space that aligns well with pre-trained Vicuna. We also build an off-the-shelf visual grounding pipeline to explore the fine-grained relation between different visual objects and modalities.

BuboGPT: Training Procedure

BuboGPT connects different modality Q-Former with pre-trained large language model Vicuna, using a simple projection matrix. We consider a two-stage instruction-tuning procedure:

Stage 1: Single-modal Pre-training. We train the corresponding modality Q-Former and linear projection layer on a large number of modality-text paired data.
Stage 2: Multi-Modal Instruct Tuning.
We curate a high-quality multi-modal instruction-following dataset to fine tune only the linear projection layer:
- Image-Text: We employ two previously published datasets from MiniGPT-4 and LLaVa for visual instruct tuning.
- Audio-Text: We build a series of expressive and descriptive data to facilitate this process based on Clotho dataset.
- Audio-Image-Text: We build <audio, image, text> pairs to act as triple-modality instruction tuning dataset based on VGGSS dataset and further introduce negative set to enhance our model.

Examples on Fine-grained Visual Understanding

We first consider using a single image as input for fine-grained visual understanding with grounding. As the exmaples shown, the model can accurately associate textural words or phrases with image regions in various scenarios with different complexities.

Examples on Audio Understanding

When a single audio clip is provided for audio understanding, BuboGPT gives informative descriptions covering nearly all acoustic parts included, even when some audio fragments are too short for humans to notice, see examples for details.

Examples on Aligned audio-image understanding

We show that BuboGPT can perform sound localization with a matched audio-image pair provided, which gives a perfect example for aligned audio-image understanding, see examples for details.

Examples on Arbitrary audio-image understanding

The BuboGPT can also tell whether the image and audio are relevant to each other and generate high-quality response for arbitrary audio-image understanding, see examples for details.

BibTeX


  @article{zhao2023bubogpt,
    author      = {Yang Zhao and Zhijie Lin and Daquan Zhou and Zilong Huang and Jiashi Feng and Bingyi Kang},
    title       = {BuboGPT: Enabling Visual Grounding in Multi-Modal LLMs},
    publisher   = {arXiv:2307.08581},
    year        = {2023}
  }