arXiv subjects

Piotr Koniusz

Publications and source records attributed to Piotr Koniusz.

At least 37 records · Page 2Linked to original sources

arXiv2025

Math Blind: Failures in Diagram Understanding Undermine Reasoning in MLLMs

Diagrams represent a form of visual language that encodes abstract concepts and relationships through structured symbols and their spatial arrangements. Unlike natural images, they are inherently symbolic, and entirely artificial. They thus pose unique challenges for Multimodal Large Language Models (MLLMs) distinct from natural image processing. Recent studies have shown that MLLMs often exhibit flawed reasoning and hallucinations when handling diagram inputs. We investigate here whether these limitations stem from shortcomings in the models' ability to interpret diagrams themselves. To this end, we develop a diagnostic test suite that isolates perception from reasoning. Our systematic evaluation reveals that MLLMs perform poorly on basic perceptual tasks, e.g., shape classification, object counting, relationship identification, and object grounding, with near-zero accuracy on fine-grained grounding. Further analysis shows that weak diagram perception leads to "blind faith in text", where models rely on textual shortcuts rather than visual understanding (that is, they are Math Blind). We hypothesize that enabling models to capture the inherent structural properties of diagrams, represented as graphs of primitives and their interrelationships, is essential for improving diagram understanding. Experiments with 7B and 32B MLLMs validate this assumption, with models trained on such representations achieving a +79% gain on the grounding task. Crucially, these gains transfer to reasoning, achieving 3-4% cross-suite improvements on three public benchmarks even without additional chain-of-thought reasoning data. Our findings demonstrate that low-level perception supports faithful high-level reasoning in mathematical MLLMs. We provide both methodological frameworks and empirical evidence to guide future research in this direction.

cs.CV↗

arXiv2025

Point-PNG: Conditional Pseudo-Negatives Generation for Point Cloud Pre-Training

We propose Point-PNG, a novel self-supervised learning framework that generates conditional pseudo-negatives in the latent space to learn point cloud representations that are both discriminative and transformation-sensitive. Conventional self-supervised learning methods focus on achieving invariance, discarding transformation-specific information. Recent approaches incorporate transformation sensitivity by explicitly modeling relationships between original and transformed inputs. However, they often suffer from an invariant-collapse phenomenon, where the predictor degenerates into identity mappings, resulting in latent representations with limited variation across transformations. To address this, we propose Point-PNG that explicitly penalizes invariant collapse through pseudo-negatives generation, enabling the network to capture richer transformation cues while preserving discriminative representations. To this end, we introduce a parametric network, COnditional Pseudo-Negatives Embedding (COPE), which learns localized displacements induced by transformations within the latent space. A key challenge arises when jointly training COPE with the MAE, as it tends to converge to trivial identity mappings. To overcome this, we design a loss function based on pseudo-negatives conditioned on the transformation, which penalizes such trivial invariant solutions and enforces meaningful representation learning. We validate Point-PNG on shape classification and relative pose estimation tasks, showing competitive performance on ModelNet40 and ScanObjectNN under challenging evaluation protocols, and achieving superior accuracy in relative pose estimation compared to supervised baselines.

cs.CV↗

arXiv2025

MMT-ARD: Multimodal Multi-Teacher Adversarial Distillation for Robust Vision-Language Models

Vision-Language Models (VLMs) are increasingly deployed in safety-critical applications, making their adversarial robustness a crucial concern. While adversarial knowledge distillation has shown promise in transferring robustness from teacher to student models, traditional single-teacher approaches suffer from limited knowledge diversity, slow convergence, and difficulty in balancing robustness and accuracy. To address these challenges, we propose MMT-ARD: a Multimodal Multi-Teacher Adversarial Robust Distillation framework. Our key innovation is a dual-teacher knowledge fusion architecture that collaboratively optimizes clean feature preservation and robust feature enhancement. To better handle challenging adversarial examples, we introduce a dynamic weight allocation strategy based on teacher confidence, enabling adaptive focus on harder samples. Moreover, to mitigate bias among teachers, we design an adaptive sigmoid-based weighting function that balances the strength of knowledge transfer across modalities. Extensive experiments on ImageNet and zero-shot benchmarks demonstrate that MMT-ARD improves robust accuracy by +4.32% and zero-shot accuracy by +3.5% on the ViT-B-32 model, while achieving a 2.3x increase in training efficiency over traditional single-teacher methods. These results highlight the effectiveness and scalability of MMT-ARD in enhancing the adversarial robustness of multimodal large models. Our codes are available at https://github.com/itsnotacie/MMT-ARD.

cs.CV↗

arXiv2025

Learning Time in Static Classifiers

Real-world visual data rarely presents as isolated, static instances. Instead, it often evolves gradually over time through variations in pose, lighting, object state, or scene context. However, conventional classifiers are typically trained under the assumption of temporal independence, limiting their ability to capture such dynamics. We propose a simple yet effective framework that equips standard feedforward classifiers with temporal reasoning, all without modifying model architectures or introducing recurrent modules. At the heart of our approach is a novel Support-Exemplar-Query (SEQ) learning paradigm, which structures training data into temporally coherent trajectories. These trajectories enable the model to learn class-specific temporal prototypes and align prediction sequences via a differentiable soft-DTW loss. A multi-term objective further promotes semantic consistency and temporal smoothness. By interpreting input sequences as evolving feature trajectories, our method introduces a strong temporal inductive bias through loss design alone. This proves highly effective in both static and temporal tasks: it enhances performance on fine-grained and ultra-fine-grained image classification, and delivers precise, temporally consistent predictions in video anomaly detection. Despite its simplicity, our approach bridges static and temporal learning in a modular and data-efficient manner, requiring only a simple classifier on top of pre-extracted features.

cs.CV↗

arXiv2025

Feature Hallucination for Self-supervised Action Recognition

Understanding human actions in videos requires more than raw pixel analysis; it relies on high-level semantic reasoning and effective integration of multimodal features. We propose a deep translational action recognition framework that enhances recognition accuracy by jointly predicting action concepts and auxiliary features from RGB video frames. At test time, hallucination streams infer missing cues, enriching feature representations without increasing computational overhead. To focus on action-relevant regions beyond raw pixels, we introduce two novel domain-specific descriptors. Object Detection Features (ODF) aggregate outputs from multiple object detectors to capture contextual cues, while Saliency Detection Features (SDF) highlight spatial and intensity patterns crucial for action recognition. Our framework seamlessly integrates these descriptors with auxiliary modalities such as optical flow, Improved Dense Trajectories, skeleton data, and audio cues. It remains compatible with state-of-the-art architectures, including I3D, AssembleNet, Video Transformer Network, FASTER, and recent models like VideoMAE V2 and InternVideo2. To handle uncertainty in auxiliary features, we incorporate aleatoric uncertainty modeling in the hallucination step and introduce a robust loss function to mitigate feature noise. Our multimodal self-supervised action recognition framework achieves state-of-the-art performance on multiple benchmarks, including Kinetics-400, Kinetics-600, and Something-Something V2, demonstrating its effectiveness in capturing fine-grained action dynamics.

cs.CV↗

arXiv2025

Noise Consistency Regularization for Improved Subject-Driven Image Synthesis

Fine-tuning Stable Diffusion enables subject-driven image synthesis by adapting the model to generate images containing specific subjects. However, existing fine-tuning methods suffer from two key issues: underfitting, where the model fails to reliably capture subject identity, and overfitting, where it memorizes the subject image and reduces background diversity. To address these challenges, we propose two auxiliary consistency losses for diffusion fine-tuning. First, a prior consistency regularization loss ensures that the predicted diffusion noise for prior (non-subject) images remains consistent with that of the pretrained model, improving fidelity. Second, a subject consistency regularization loss enhances the fine-tuned model's robustness to multiplicative noise modulated latent code, helping to preserve subject identity while improving diversity. Our experimental results demonstrate that incorporating these losses into fine-tuning not only preserves subject identity but also enhances image diversity, outperforming DreamBooth in terms of CLIP scores, background variation, and overall visual quality.

cs.GR↗

arXiv2025

CHAIN: Enhancing Generalization in Data-Efficient GANs via lipsCHitz continuity constrAIned Normalization

Generative Adversarial Networks (GANs) significantly advanced image generation but their performance heavily depends on abundant training data. In scenarios with limited data, GANs often struggle with discriminator overfitting and unstable training. Batch Normalization (BN), despite being known for enhancing generalization and training stability, has rarely been used in the discriminator of Data-Efficient GANs. Our work addresses this gap by identifying a critical flaw in BN: the tendency for gradient explosion during the centering and scaling steps. To tackle this issue, we present CHAIN (lipsCHitz continuity constrAIned Normalization), which replaces the conventional centering step with zero-mean regularization and integrates a Lipschitz continuity constraint in the scaling step. CHAIN further enhances GAN training by adaptively interpolating the normalized and unnormalized features, effectively avoiding discriminator overfitting. Our theoretical analyses firmly establishes CHAIN's effectiveness in reducing gradients in latent features and weights, improving stability and generalization in GAN training. Empirical evidence supports our theory. CHAIN achieves state-of-the-art results in data-limited scenarios on CIFAR-10/100, ImageNet, five low-shot and seven high-resolution few-shot image datasets. Code: https://github.com/MaxwellYaoNi/CHAIN

cs.LG↗

arXiv2025

Learnable Expansion of Graph Operators for Multi-Modal Feature Fusion

In computer vision tasks, features often come from diverse representations, domains (e.g., indoor and outdoor), and modalities (e.g., text, images, and videos). Effectively fusing these features is essential for robust performance, especially with the availability of powerful pre-trained models like vision-language models. However, common fusion methods, such as concatenation, element-wise operations, and non-linear techniques, often fail to capture structural relationships, deep feature interactions, and suffer from inefficiency or misalignment of features across domains or modalities. In this paper, we shift from high-dimensional feature space to a lower-dimensional, interpretable graph space by constructing relationship graphs that encode feature relationships at different levels, e.g., clip, frame, patch, token, etc. To capture deeper interactions, we expand graphs through iterative graph relationship updates and introduce a learnable graph fusion operator to integrate these expanded relationships for more effective fusion. Our approach is relationship-centric, operates in a homogeneous space, and is mathematically principled, resembling element-wise relationship score aggregation via multilinear polynomials. We demonstrate the effectiveness of our graph-based fusion method on video anomaly detection, showing strong performance across multi-representational, multi-modal, and multi-domain feature fusion tasks.

cs.CV↗

arXiv2025

Understanding and Mitigating Hyperbolic Dimensional Collapse in Graph Contrastive Learning

Learning generalizable self-supervised graph representations for downstream tasks is challenging. To this end, Contrastive Learning (CL) has emerged as a leading approach. The embeddings of CL are arranged on a hypersphere where similarity is measured by the cosine distance. However, many real-world graphs, especially of hierarchical nature, cannot be embedded well in the Euclidean space. Although the hyperbolic embedding is suitable for hierarchical representation learning, naively applying CL to the hyperbolic space may result in the so-called dimension collapse, i.e., features will concentrate mostly within few density regions, leading to poor utilization of the whole feature space. Thus, we propose a novel contrastive learning framework to learn high-quality graph embeddings in hyperbolic space. Specifically, we design the alignment metric that effectively captures the hierarchical data-invariant information, as well as we propose a substitute of the uniformity metric to prevent the so-called dimensional collapse. We show that in the hyperbolic space one has to address the leaf- and height-level uniformity related to properties of trees. In the ambient space of the hyperbolic manifold these notions translate into imposing an isotropic ring density towards boundaries of Poincaré ball. Our experiments support the efficacy of our method.

cs.LG↗

arXiv2025

Graph Self-Supervised Learning with Learnable Structural and Positional Encodings

Traditional Graph Self-Supervised Learning (GSSL) struggles to capture complex structural properties well. This limitation stems from two main factors: (1) the inadequacy of conventional Graph Neural Networks (GNNs) in representing sophisticated topological features, and (2) the focus of self-supervised learning solely on final graph representations. To address these issues, we introduce \emph{GenHopNet}, a GNN framework that integrates a $k$-hop message-passing scheme, enhancing its ability to capture local structural information without explicit substructure extraction. We theoretically demonstrate that \emph{GenHopNet} surpasses the expressiveness of the classical Weisfeiler-Lehman (WL) test for graph isomorphism. Furthermore, we propose a structural- and positional-aware GSSL framework that incorporates topological information throughout the learning process. This approach enables the learning of representations that are both sensitive to graph topology and invariant to specific structural and feature augmentations. Comprehensive experiments on graph classification datasets, including those designed to test structural sensitivity, show that our method consistently outperforms the existing approaches and maintains computational efficiency. Our work significantly advances GSSL's capability in distinguishing graphs with similar local structures but different global topologies.

cs.LG↗

arXiv2025

Open Eyes, Then Reason: Fine-grained Visual Mathematical Understanding in MLLMs

Current multimodal large language models (MLLMs) often underperform on mathematical problem-solving tasks that require fine-grained visual understanding. The limitation is largely attributable to inadequate perception of geometric primitives during image-level contrastive pre-training (e.g., CLIP). While recent efforts to improve math MLLMs have focused on scaling up mathematical visual instruction datasets and employing stronger LLM backbones, they often overlook persistent errors in visual recognition. In this paper, we systematically evaluate the visual grounding capabilities of state-of-the-art MLLMs and reveal a significant negative correlation between visual grounding accuracy and problem-solving performance, underscoring the critical role of fine-grained visual understanding. Notably, advanced models like GPT-4o exhibit a 70% error rate when identifying geometric entities, highlighting that this remains a key bottleneck in visual mathematical reasoning. To address this, we propose a novel approach, SVE-Math (Selective Vision-Enhanced Mathematical MLLM), featuring a geometric-grounded vision encoder and a feature router that dynamically adjusts the contribution of hierarchical visual feature maps. Our model recognizes accurate visual primitives and generates precise visual prompts tailored to the language model's reasoning needs. In experiments, SVE-Math-Qwen2.5-7B outperforms other 7B models by 15% on MathVerse and is compatible with GPT-4V on MathVista. Despite being trained on smaller datasets, SVE-Math-7B achieves competitive performance on GeoQA, rivaling models trained on significantly larger datasets. Our findings emphasize the importance of incorporating fine-grained visual understanding into MLLMs and provide a promising direction for future research.

cs.CV↗

arXiv2024

When Spatial meets Temporal in Action Recognition

Video action recognition has made significant strides, but challenges remain in effectively using both spatial and temporal information. While existing methods often focus on either spatial features (e.g., object appearance) or temporal dynamics (e.g., motion), they rarely address the need for a comprehensive integration of both. Capturing the rich temporal evolution of video frames, while preserving their spatial details, is crucial for improving accuracy. In this paper, we introduce the Temporal Integration and Motion Enhancement (TIME) layer, a novel preprocessing technique designed to incorporate temporal information. The TIME layer generates new video frames by rearranging the original sequence, preserving temporal order while embedding $N^2$ temporally evolving frames into a single spatial grid of size $N \times N$. This transformation creates new frames that balance both spatial and temporal information, making them compatible with existing video models. When $N=1$, the layer captures rich spatial details, similar to existing methods. As $N$ increases ($N\geq2$), temporal information becomes more prominent, while the spatial information decreases to ensure compatibility with model inputs. We demonstrate the effectiveness of the TIME layer by integrating it into popular action recognition models, such as ResNet-50, Vision Transformer, and Video Masked Autoencoders, for both RGB and depth video data. Our experiments show that the TIME layer enhances recognition accuracy, offering valuable insights for video processing tasks.

cs.CV↗

arXiv2024

Inductive Graph Few-shot Class Incremental Learning

Node classification with Graph Neural Networks (GNN) under a fixed set of labels is well known in contrast to Graph Few-Shot Class Incremental Learning (GFSCIL), which involves learning a GNN classifier as graph nodes and classes growing over time sporadically. We introduce inductive GFSCIL that continually learns novel classes with newly emerging nodes while maintaining performance on old classes without accessing previous data. This addresses the practical concern of transductive GFSCIL, which requires storing the entire graph with historical data. Compared to the transductive GFSCIL, the inductive setting exacerbates catastrophic forgetting due to inaccessible previous data during incremental training, in addition to overfitting issue caused by label sparsity. Thus, we propose a novel method, called Topology-based class Augmentation and Prototype calibration (TAP). To be specific, it first creates a triple-branch multi-topology class augmentation method to enhance model generalization ability. As each incremental session receives a disjoint subgraph with nodes of novel classes, the multi-topology class augmentation method helps replicate such a setting in the base session to boost backbone versatility. In incremental learning, given the limited number of novel class samples, we propose an iterative prototype calibration to improve the separation of class prototypes. Furthermore, as backbone fine-tuning poses the feature distribution drift, prototypes of old classes start failing over time, we propose the prototype shift method for old classes to compensate for the drift. We showcase the proposed method on four datasets.

cs.LG↗

arXiv2024

Less is More: Extreme Gradient Boost Rank-1 Adaption for Efficient Finetuning of LLMs

Fine-tuning Large Language Models (LLMs) has become a crucial technique for adapting pre-trained models to downstream tasks. However, the enormous size of LLMs poses significant challenges in terms of computational complexity and resource requirements. Low-Rank Adaptation (LoRA) has emerged as a promising solution. However, there exists a gap between the practical performance of low-rank adaptations and its theoretical optimum. In this work, we propose eXtreme Gradient Boosting LoRA (XGBLoRA), a novel framework that bridges this gap by leveraging the power of ensemble learning. Inspired by gradient boosting, XGBLoRA iteratively learns and merges a sequence of LoRA adaptations to refine model predictions. It achieves better performance than the standard LoRA, while enjoying the computational efficiency of rank-1 adaptations. We provide theoretical analysis to show the convergence and optimality of our approach, and conduct extensive experiments on a range of natural language processing tasks. The results demonstrate that XGBLoRA consistently outperforms standard LoRA and achieves performance comparable to full fine-tuning with significantly fewer trainable parameters. This work advances parameter-efficient fine-tuning for LLMs, and offers a promising solution for adapting LLMs to downstream tasks while optimizing performance and efficiency.

cs.CL↗

arXiv2024

Motion meets Attention: Video Motion Prompts

Videos contain rich spatio-temporal information. Traditional methods for extracting motion, used in tasks such as action recognition, often rely on visual contents rather than precise motion features. This phenomenon is referred to as 'blind motion extraction' behavior, which proves inefficient in capturing motions of interest due to a lack of motion-guided cues. Recently, attention mechanisms have enhanced many computer vision tasks by effectively highlighting salient visual areas. Inspired by this, we propose a modified Sigmoid function with learnable slope and shift parameters as an attention mechanism to modulate motion signals from frame differencing maps. This approach generates a sequence of attention maps that enhance the processing of motion-related video content. To ensure temporal continuity and smoothness of the attention maps, we apply pair-wise temporal attention variation regularization to remove unwanted motions (e.g., noise) while preserving important ones. We then perform Hadamard product between each pair of attention maps and the original video frames to highlight the evolving motions of interest over time. These highlighted motions, termed video motion prompts, are subsequently used as inputs to the model instead of the original video frames. We formalize this process as a motion prompt layer and incorporate the regularization term into the loss function to learn better motion prompts. This layer serves as an adapter between the model and the video data, bridging the gap between traditional 'blind motion extraction' and the extraction of relevant motions of interest. We show that our lightweight, plug-and-play motion prompt layer seamlessly integrates into models like SlowFast, X3D, and TimeSformer, enhancing performance on benchmarks such as FineGym and MPII Cooking 2.

cs.CV↗

arXiv2024

OpenKD: Opening Prompt Diversity for Zero- and Few-shot Keypoint Detection

Exploiting the foundation models (e.g., CLIP) to build a versatile keypoint detector has gained increasing attention. Most existing models accept either the text prompt (e.g., ``the nose of a cat''), or the visual prompt (e.g., support image with keypoint annotations), to detect the corresponding keypoints in query image, thereby, exhibiting either zero-shot or few-shot detection ability. However, the research on taking multimodal prompt is still underexplored, and the prompt diversity in semantics and language is far from opened. For example, how to handle unseen text prompts for novel keypoint detection and the diverse text prompts like ``Can you detect the nose and ears of a cat?'' In this work, we open the prompt diversity from three aspects: modality, semantics (seen v.s. unseen), and language, to enable a more generalized zero- and few-shot keypoint detection (Z-FSKD). We propose a novel OpenKD model which leverages multimodal prototype set to support both visual and textual prompting. Further, to infer the keypoint location of unseen texts, we add the auxiliary keypoints and texts interpolated from visual and textual domains into training, which improves the spatial reasoning of our model and significantly enhances zero-shot novel keypoint detection. We also found large language model (LLM) is a good parser, which achieves over 96% accuracy to parse keypoints from texts. With LLM, OpenKD can handle diverse text prompts. Experimental results show that our method achieves state-of-the-art performance on Z-FSKD and initiates new ways to deal with unseen text and diverse texts. The source code and data are available at https://github.com/AlanLuSun/OpenKD.

cs.CV↗

arXiv2024

Adaptive Multi-head Contrastive Learning

In contrastive learning, two views of an original image, generated by different augmentations, are considered a positive pair, and their similarity is required to be high. Similarly, two views of distinct images form a negative pair, with encouraged low similarity. Typically, a single similarity measure, provided by a lone projection head, evaluates positive and negative sample pairs. However, due to diverse augmentation strategies and varying intra-sample similarity, views from the same image may not always be similar. Additionally, owing to inter-sample similarity, views from different images may be more akin than those from the same image. Consequently, enforcing high similarity for positive pairs and low similarity for negative pairs may be unattainable, and in some cases, such enforcement could detrimentally impact performance. To address this challenge, we propose using multiple projection heads, each producing a distinct set of features. Our pre-training loss function emerges from a solution to the maximum likelihood estimation over head-wise posterior distributions of positive samples given observations. This loss incorporates the similarity measure over positive and negative pairs, each re-weighted by an individual adaptive temperature, regulated to prevent ill solutions. Our approach, Adaptive Multi-Head Contrastive Learning (AMCL), can be applied to and experimentally enhances several popular contrastive learning methods such as SimCLR, MoCo, and Barlow Twins. The improvement remains consistent across various backbones and linear probing epochs, and becomes more significant when employing multiple augmentation methods.

cs.CV↗

arXiv2024

Pre-training with Random Orthogonal Projection Image Modeling

Masked Image Modeling (MIM) is a powerful self-supervised strategy for visual pre-training without the use of labels. MIM applies random crops to input images, processes them with an encoder, and then recovers the masked inputs with a decoder, which encourages the network to capture and learn structural information about objects and scenes. The intermediate feature representations obtained from MIM are suitable for fine-tuning on downstream tasks. In this paper, we propose an Image Modeling framework based on random orthogonal projection instead of binary masking as in MIM. Our proposed Random Orthogonal Projection Image Modeling (ROPIM) reduces spatially-wise token information under guaranteed bound on the noise variance and can be considered as masking entire spatial image area under locally varying masking degrees. Since ROPIM uses a random subspace for the projection that realizes the masking step, the readily available complement of the subspace can be used during unmasking to promote recovery of removed information. In this paper, we show that using random orthogonal projection leads to superior performance compared to crop-based masking. We demonstrate state-of-the-art results on several popular benchmarks.

cs.CV↗