Multimodal Fusion: MIVPG’s Hierarchical MIL Approach for Multi-Image Samples | HackerNoon

Table of Links

Abstract and 1 Introduction

2. Related Work

   2.1. Multimodal Learning

   2.2. Multiple Instance Learning

3. Methodology

   3.1. Preliminaries and Notations

   3.2. Relations between Attention-based VPG and MIL

   3.3. MIVPG for Multiple Visual Inputs

   3.4. Unveiling Instance Correlation in MIVPG for Enhanced Multi-instance Scenarios

4. Experiments and 4.1. General Setup

   4.2. Scenario 1: Samples with Single Image

   4.3. Scenario 2: Samples with Multiple Images, with Each Image as a General Embedding

   4.4. Scenario 3: Samples with Multiple Images, with Each Image Having Multiple Patches to be Considered and 4.5. Case Study

5. Conclusion and References

Supplementary Material

A. Detailed Architecture of QFormer

B. Proof of Proposition

C. More Experiments

3.3. MIVPG for Multiple Visual Inputs

When a sample comprises multiple images, it is imperative to consider MIL feature aggregation from different perspectives. In the context of individual images, each image can be treated as a ’bag,’ and each patch within the image as an ’instance.’ From the sample’s perspective, each sample can also be regarded as a ’bag,’ with each image within the sample as an ’instance.’ When a sample contains only a single image, we can focus primarily on the former perspective since the latter perspective involves a single instance per bag. However, in a more general context, it is essential to adopt a hierarchical approach when considering the utilization of MIL for feature aggregation. Without loss of generality, we now consider the input of the MIVPG to be a bag B containing multiple instances. Hence, the cross-attention can be expressed as Attention(Q = q, K = B, V = B).

:::info
Authors:

(1) Wenliang Zhong, The University of Texas at Arlington ([email protected]);

(2) Wenyi Wu, Amazon ([email protected]);

(3) Qi Li, Amazon ([email protected]);

(4) Rob Barton, Amazon ([email protected]);

(5) Boxin Du, Amazon ([email protected]);

(6) Shioulin Sam, Amazon ([email protected]);

(7) Karim Bouyarmane, Amazon ([email protected]);

(8) Ismail Tutar, Amazon ([email protected]);

(9) Junzhou Huang, The University of Texas at Arlington ([email protected]).

:::

:::info
This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license.

:::