Magic-MM-Embedding: Towards Visual-Token-Efficient Universal Multimodal Embedding with MLLMs

Multimodal Large Language Models (MLLMs) have shown immense promise in universal multimodal retrieval, which aims to find relevant items of various modalities for a given query. But their practical application is often hindered by the substantial computational cost incurred from processing a large number of tokens from visual inputs. In this paper, we propose Magic-MM-Embedding, a series of novel models that achieve both high efficiency and state-of-the-art performance in universal multimodal embedding. Our approach is built on two synergistic pillars: (1) a highly efficient MLLM architecture incorporating visual token compression to drastically reduce inference latency and memory footprint, and (2) a multi-stage progressive training strategy designed to not only recover but significantly boost performance. This coarse-to-fine training paradigm begins with extensive continue pretraining to restore multimodal understanding and generation capabilities, progresses to large-scale contrastive pretraining and hard negative mining to enhance discriminative power, and culminates in a task-aware fine-tuning stage guided by an MLLM-as-a-Judge for precise data curation. Comprehensive experiments show that our model outperforms existing methods by a large margin while being more inference-efficient.

Imbalance Knowledge-Driven Multi-modal Network for Land-Cover Semantic Segmentation Using Images and LiDAR Point Clouds

Despite the good results that have been achieved in unimodal segmentation, the inherent limitations of individual data increase the difficulty of achieving breakthroughs in performance. For that reason, multi-modal learning is increasingly being explored within the field of remote sensing. The present multi-modal methods usually map high-dimensional features to low-dimensional spaces as a preprocess before feature extraction to address the nonnegligible domain gap, which inevitably leads to information loss. To address this issue, in this paper we present our novel Imbalance Knowledge-Driven Multi-modal Network (IKD-Net) to extract features from raw multi-modal heterogeneous data directly. IKD-Net is capable of mining imbalance information across modalities while utilizing a strong modal to drive the feature map refinement of the weaker ones in the global and categorical perspectives by way of two sophisticated plug-and-play modules: the Global Knowledge-Guided (GKG) and Class Knowledge-Guided (CKG) gated modules. The whole network then is optimized using a holistic loss function. While we were developing IKD-Net, we also established a new dataset called the National Agriculture Imagery Program and 3D Elevation Program Combined dataset in California (N3C-California), which provides a particular benchmark for multi-modal joint segmentation tasks. In our experiments, IKD-Net outperformed the benchmarks and state-of-the-art methods both in the N3C-California and the small-scale ISPRS Vaihingen dataset. IKD-Net has been ranked first on the real-time leaderboard for the GRSS DFC 2018 challenge evaluation until this paper's submission.