Calibrated Multimodal Representation Learning with Missing Modalities
Multimodal representation learning under partial-modality settings
✨ Overview
CalMRL is a multimodal representation learning framework designed for alignment calibration when some modalities are missing. CalMRL combines two complementary goals:
- Cross-modal alignment for robust shared representations
- Missing-modality calibration through posterior inference and learned generative parameters
🎯 Key Features
🔄 Partial-Modality Learning
- Handles missing video, audio, text, or subtitle signals
- Supports posterior-based feature completion with learned modality-specific parameters
🎯 Multimodal Retrieval
- Joint training over text-video, text-audio, text-video-audio, and subtitle-aware setups
- Config-driven recipes for pretraining, finetuning, and evaluation
🧠 Feature Calibration
- Uses latent posterior inference for modality completion
- Includes a warmup pipeline to estimate
W,mu, andlog_sigma
🏗️ Architecture
The current codebase is organized around three main stages:
- 🔧 Multimodal Encoding: Video, audio, text, and subtitle features are extracted with VAST-style encoders.
- 🧮 Representation Calibration: Shared embeddings are aligned while latent posterior inference estimates missing information.
- 🔄 Downstream Evaluation: Retrieval and other tasks are executed through a unified config-driven pipeline.
Citation
If this project is useful for your research, you can cite the work as:
@article{liu2025calibrated,
title={Calibrated Multimodal Representation Learning with Missing Modalities},
author={Liu, Xiaohao and Xia, Xiaobo and Wei, Jiaheng and Yang, Shuo and Su, Xiu and Ng, See-Kiong and Chua, Tat-Seng},
journal={arXiv preprint arXiv:2511.12034},
year={2025}
}