MM-Mixing: Multi-Modal Mixing Alignment for 3D Understanding

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Jiaze, Wang, Yi, Guo, Ziyu, Zhang, Renrui, Zhou, Donghao, Chen, Guangyong, Liu, Anfeng, Heng, Pheng-Ann
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929462624387072
author Wang, Jiaze
Wang, Yi
Guo, Ziyu
Zhang, Renrui
Zhou, Donghao
Chen, Guangyong
Liu, Anfeng
Heng, Pheng-Ann
author_facet Wang, Jiaze
Wang, Yi
Guo, Ziyu
Zhang, Renrui
Zhou, Donghao
Chen, Guangyong
Liu, Anfeng
Heng, Pheng-Ann
contents We introduce MM-Mixing, a multi-modal mixing alignment framework for 3D understanding. MM-Mixing applies mixing-based methods to multi-modal data, preserving and optimizing cross-modal connections while enhancing diversity and improving alignment across modalities. Our proposed two-stage training pipeline combines feature-level and input-level mixing to optimize the 3D encoder. The first stage employs feature-level mixing with contrastive learning to align 3D features with their corresponding modalities. The second stage incorporates both feature-level and input-level mixing, introducing mixed point cloud inputs to further refine 3D feature representations. MM-Mixing enhances intermodality relationships, promotes generalization, and ensures feature consistency while providing diverse and realistic training samples. We demonstrate that MM-Mixing significantly improves baseline performance across various learning scenarios, including zero-shot 3D classification, linear probing 3D classification, and cross-modal 3D shape retrieval. Notably, we improved the zero-shot classification accuracy on ScanObjectNN from 51.3% to 61.9%, and on Objaverse-LVIS from 46.8% to 51.4%. Our findings highlight the potential of multi-modal mixing-based alignment to significantly advance 3D object recognition and understanding while remaining straightforward to implement and integrate into existing frameworks.
format Preprint
id arxiv_https___arxiv_org_abs_2405_18523
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle MM-Mixing: Multi-Modal Mixing Alignment for 3D Understanding
Wang, Jiaze
Wang, Yi
Guo, Ziyu
Zhang, Renrui
Zhou, Donghao
Chen, Guangyong
Liu, Anfeng
Heng, Pheng-Ann
Computer Vision and Pattern Recognition
Artificial Intelligence
We introduce MM-Mixing, a multi-modal mixing alignment framework for 3D understanding. MM-Mixing applies mixing-based methods to multi-modal data, preserving and optimizing cross-modal connections while enhancing diversity and improving alignment across modalities. Our proposed two-stage training pipeline combines feature-level and input-level mixing to optimize the 3D encoder. The first stage employs feature-level mixing with contrastive learning to align 3D features with their corresponding modalities. The second stage incorporates both feature-level and input-level mixing, introducing mixed point cloud inputs to further refine 3D feature representations. MM-Mixing enhances intermodality relationships, promotes generalization, and ensures feature consistency while providing diverse and realistic training samples. We demonstrate that MM-Mixing significantly improves baseline performance across various learning scenarios, including zero-shot 3D classification, linear probing 3D classification, and cross-modal 3D shape retrieval. Notably, we improved the zero-shot classification accuracy on ScanObjectNN from 51.3% to 61.9%, and on Objaverse-LVIS from 46.8% to 51.4%. Our findings highlight the potential of multi-modal mixing-based alignment to significantly advance 3D object recognition and understanding while remaining straightforward to implement and integrate into existing frameworks.
title MM-Mixing: Multi-Modal Mixing Alignment for 3D Understanding
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2405.18523