Open-vocabulary semantic segmentation (OVSS) employs pixel-level vision-language alignment to associate category-related prompts with corresponding pixels. A key challenge is enhancing the multimodal dense prediction capability, specifically this pixel-level multimodal alignment. Although existing methods achieve promising results by leveraging CLIP’s vision-language alignment, they rarely investigate the performance boundaries of CLIP for dense prediction from an interpretability mechanisms perspective. In this work, we systematically investigate CLIP's internal mechanisms and identify a critical phenomenon: analogous to human distraction, CLIP diverts significant attention resources from target regions to irrelevant tokens. Our analysis reveals that these tokens arise from dimension-specific over-activation; filtering them enhances CLIP's dense prediction performance. Consequently, we propose Refocusing CLIP (RF-CLIP), a training-free approach that emulates human distraction-refocusing behavior to redirect attention from distraction tokens back to target regions, thereby refining CLIP's multimodal alignment granularity. Our method achieves SOTA performance on eight benchmarks while maintaining high inference efficiency.

3D Visual Grounding (3DVG) aims to localize the referent of natural language referring expressions through two core tasks: Referring Expression Comprehension (3DREC) and Segmentation (3DRES). While existing methods achieve high accuracy in simple, single-object scenes, they suffer from severe performance degradation in complex, multi-object scenes that are common in real-world settings, hindering practical deployment. Existing methods face two key challenges in complex, multi-object scenes: inadequate parsing of implicit localization cues critical for disambiguating visually similar objects, and ineffective suppression of dynamic spatial interference from co-occurring objects, resulting in degraded grounding accuracy. To address these challenges, we propose PC-CrossDiff, a unified dual-task framework with a dual-level cross-modal differential attention architecture for 3DREC and 3DRES. Specifically, the framework introduces: (i) Point-Level Differential Attention (PLDA) modules that apply bidirectional differential attention between text and point clouds, adaptively extracting implicit localization cues via learnable weights to improve discriminative representation; (ii) Cluster-Level Differential Attention (CLDA) modules that establish a hierarchical attention mechanism to adaptively enhance localization-relevant spatial relationships while suppressing ambiguous or irrelevant spatial relations through a localization-aware differential attention block. To address the scale disparity and conflicting gradients in joint 3DREC–3DRES training, we propose L_DGTL, a unified loss function that explicitly reduces multi-task crosstalk and enables effective parameter sharing across tasks. Our method achieves state-of-the-art performance on the ScanRefer, NR3D, and SR3D benchmarks. Notably, on the Implicit subsets of ScanRefer, it improves the [email protected] score by +10.16% for the 3DREC task, highlighting its strong ability to parse implicit spatial cues.

Domain generalization (DG) and domain adaptation (DA) for 3D semantic segmentation enable the model to maintain high performance while avoiding labor-intensive and time-consuming annotation of target-domain data. However, under adverse weather conditions, the injection of spatial noise will affect the reflectivity of LiDAR point clouds, exacerbate domain distribution discrepancies, and degrade the generalization ability of the model. Current methods mainly rely on sparse convolution-based architecture. Due to its limited receptive field, the model captures varying local geometric information when dealing with point clouds of different sparsities, thereby limiting its transferability. To this end, we propose BeyondSparse, a novel cross-domain 3D semantic segmentation method under adverse weather that incorporates a state-space model into a 3D sparse convolution-based architecture, sequentially modeling all features to learn domain-invariant representations. This method consists of two main components: domain feature decoupling and Mamba-based encoder. The former performs feature disentanglement before sequential modeling, while the latter performs global modeling on voxelized point cloud data. In addition, we introduce a token-style augmentation to capture the intrinsic properties of input data. Extensive experimental results demonstrate that our method outperforms SOTA competitors in both DG and DA tasks, for instance, achieving +4.6% and +0.8% mIoU on ``SynLiDAR to SemanticSTF''.

3D semantic segmentation serves as a fundamental component in many applications, such as autonomous driving and medical image analysis. Although recent methods have advanced the field, adapting these methods to new environments or object categories without extensive retraining remains a significant challenge. To address this, we introduce xMHashSeg, a novel training-free cross-modal LiDAR semantic segmentation framework. xMHashSeg leverages foundation models and non-parametric network to extract features from 2D images and 3D point clouds, subsequently integrating these features through hash learning. Specifically, We develop point-SANN, a novel self-adaption non-parametric network that can extract robust 3D features from raw point clouds, while 2D features are directly extracted through the foundation model DINOv2. To reconcile inconsistencies across different modals, we introduce a Hash Code Learning Module that projects all information into a common hash space, learning a consistent hash code that enhances feature integration. Additionally, depth maps are utilized as an intermediary form between 2D and 3D data to facilitate convergence during hash code learning. Our experimental results on various multi-modality datasets demonstrate that xMHashSeg outperforms zero-shot learning approaches and achieve performance close to that of unsupervised domain adaptation and test-time adaptation methods, without requiring any annotations or additional training.

Open-vocabulary semantic segmentation (OVSS) conducts pixel-level classification via text-driven alignment, where the domain discrepancy between base category training and open-vocabulary inference poses challenges in discriminative modeling of latent unseen category. To address this challenge, existing vision-language model (VLM)-based approaches demonstrate commendable performance through pre-trained multi-modal representations. However, the fundamental mechanisms of latent semantic comprehension remain underexplored, making the bottleneck for OVSS. In this work, we initiate a probing experiment to explore distribution patterns and dynamics of latent semantics in VLMs under inductive learning paradigms. Building on these insights, we propose X-Agent, an innovative OVSS framework employing latent semantic-aware ''agent'' to orchestrate cross-modal attention mechanisms, simultaneously optimizing latent semantic dynamic and amplifying its perceptibility. Extensive benchmark evaluations demonstrate that X-Agent achieves state-of-the-art performance while effectively enhancing the latent semantic saliency.