arXiv subjects

Xiying Li

Publications and source records attributed to Xiying Li.

8 recordsLinked to original sources

arXiv2026

PredErase: Training-Free Object-and-Effect Removal with Predictive Latent Guidance

Removing an object is not the same as filling its mask. Cast shadows and contact shading usually lie outside the user-provided instance mask M_obj, so a frozen Fill model that edits only that mask leaves the object's photometric footprint on nearby surfaces. Supervised removers learn this joint erasure from paired clean plates. Training-free editors freeze pretrained weights, yet most still treat M_obj as the entire editable support and steer sampling with CLIP or DINO energies that do not predict the occluded scene. We present PredErase, a training-free inference procedure on frozen FLUX.2 and I-JEPA. The method separates where Fill may rewrite pixels from what structure should occupy the hole. A contact-band expansion M_flux of M_obj exposes local residuals on the supporting plane. I-JEPA, pretrained for masked token prediction, supplies a context-conditioned hole target in representation space; sparse projected gradients align decoded Fill completions with that target inside the instance, while coordinates outside the packed support stay locked. Under instance-only masks on RemovalBench, RORD-Val, and DEFACTO-Val, PredErase improves the native FLUX.2 backbone. Supervised removers remain stronger on several full-image appearance metrics; the supported claim is training-free object-and-effect editing of frozen Fill, not replacement of paired-data erasers.

cs.CV↗

arXiv2026

Beyond Scene Priors: Fine-Grained Traffic Scene Reasoning with Benchmarking and Query-Guided Small-Object Focus

In safety-critical traffic scenarios, answering complex questions relies on minute, localized visual cues. However, standard Multimodal Large Language Models (MLLMs) tend to over-attend to backgrounds, overwhelming crucial small objects during visual-language alignment, a failure mode we term 'critical evidence dilution.' Furthermore, existing visual question answering (VQA) datasets rarely expose this flaw, as they lack large-scale, distractor-heavy evaluations that require pinpointing local evidence. To bridge this evaluation and architecture gap, we introduce the Fine-Grained Traffic Reasoning Benchmark (FGTR-Bench) and the Text-Guided Small-Object Reasoning MLLM (TSR-MLLM). FGTR-Bench comprises 40,236 single-image Multiple-Choice Questions (MCQs) created via multi-agent generation, consistency checks, and expert audits, alongside a disjoint 4,947-sample blind test split. To resolve evidence dilution, TSR-MLLM, built on Qwen3-VL-4B, uses a query-conditioned Text-Guided Small-Object Focus (TG-SOF) map. Applied once at the decoder boundary, the map adds sparse Top-K gated residuals to the most question-relevant vision slots while leaving text tokens unchanged. Together with lightweight decoder adaptation, TSR-MLLM preserves single-pass inference without external detectors or image re-encoding. Under matched settings, TSR-MLLM outperforms the strongest 4B baseline by 2.1 points on FGTR-Bench (74.1% overall), with larger gains on evidence-local tracks. Furthermore, it remains competitive on DriveQA-V (CARLA Signs) under greedy decoding without task-specific fine-tuning.

cs.CV↗

arXiv2026

IFNSO: Iteration-Free Newton-Schulz Orthogonalization

The Newton-Schulz (NS) iteration has become a key technique for orthogonalization in optimizers such as Muon and for optimization on the Stiefel manifold. Despite its effectiveness, the conventional NS iteration incurs significant computational overhead due to repeated high-dimensional matrix multiplications. To overcome these limitations, we propose Iteration-Free Newton-Schulz Orthogonalization (IFNSO), a novel framework that consolidates the traditional iterative structure into a unified and Iteration-Free formulation. By analyzing the contribution of individual matrix powers, we streamline the process by removing insignificant terms and introducing a polynomial with learnable coefficients. These coefficients are optimized to ensure both superior computational efficiency and stable convergence. Extensive experiments demonstrate that IFNSO achieves superior performance compared to existing methods. Our code is available at: https://github.com/greekinRoma/Unified_Newton_Schulz_Orthogonalization.

cs.LG↗

arXiv2025

STGBD-Net: Spatio-temporal Gradient Basis Decomposition Network for Infrared Small Target Detection

A key challenge in infrared small target detection (IRSTD) is that weak target signal responses are easily obscured by strong background clutter, frequently resulting in missed detections. While traditional gradient-based methods attempt to capture fine details, their robustness is limited by the static fusion of multi-directional gradient features. In this paper, we rethink feature fusion from the perspective of Basis Decomposition Theory and propose a novel framework that reformulates the process into an explicit and adaptive decomposition-and-reconstruction paradigm. Specifically, we introduce the Basis Decomposition Module (BDM) and its specialized variant, the Gradient Decomposition Module (GDM) for IRSTD. GDMs treat the normalized gradient features as basis vectors to reconstruct a new feature, thereby maintaining detailed structures and highlighting infrared small targets. By integrating GDMs into a lightweight three-stage U-Net, we develop two unified architectures: the Spatial Gradient Basis Decomposition Network for single-frame detection and the Spatio-temporal Gradient Basis Decomposition Network for multi-frame scenarios. Extensive experiments demonstrate that our networks achieve state-of-the-art (SOTA) performance across multiple benchmarks, offering a superior balance between detection accuracy and computational efficiency. Our codes will be made public at: https://github.com/greekinRoma/IRSTD_HC_Platform.

cs.CV↗

arXiv2025

Traffic-MLLM: Curiosity-Regularized Supervised Learning for Traffic Scenario Case-Based Reasoning

For safe and robust autonomous driving, decision-making systems must effectively leverage past experiences to handle the inherent long-tail of traffic scenarios. Case-Based Reasoning (CBR) provides a natural paradigm for this by adapting solutions from prior cases. However, in complex and dynamic traffic environments, traditional CBR methods struggle to effectively abstract and adapt knowledge under uncertainty. Meanwhile, although multimodal large language models (MLLMs) exhibit strong perceptual and linguistic capabilities, their reasoning behavior often relies on empirical pattern fitting, limiting robustness under distribution shift and long-tail scenarios. We propose Traffic-MLLM, a retrieval-free neural case modeling framework for multimodal traffic reasoning. Instead of performing explicit case retrieval at inference time, Traffic-MLLM learns a structured and generalizable case space directly during training. To support this learning process, we construct a multi-source case base by integrating dynamic traffic videos and large-scale static visual question-answering data, serving as a unified training substrate for learning structured case representations. To further improve representation quality near knowledge boundaries, we introduce a curiosity-driven refinement mechanism based on Random Network Distillation (RND), encouraging the model to internalize cross-case structural regularities rather than surface correlations. Experiments on the SUTD-TrafficQA and DriveQA benchmarks demonstrate consistent improvements in dynamic reasoning, regulatory understanding, and cross-domain transfer. Traffic-MLLM achieves 50.8% accuracy on SUTD-TrafficQA, 74.8% on the CARLA-based DriveQA split, and 83.1% on the real-world Mapillary split, indicating that representation-level case-space refinement provides an effective alternative to explicit retrieval for scalable multimodal case adaptation.

cs.CV↗

arXiv2025

Beyond Pixels: Introducing Geometric-Semantic World Priors for Video-based Embodied Models via Spatio-temporal Alignment

Achieving human-like reasoning in deep learning models for complex tasks in unknown environments remains a critical challenge in embodied intelligence. While advanced vision-language models (VLMs) excel in static scene understanding, their limitations in spatio-temporal reasoning and adaptation to dynamic, open-set tasks like task-oriented navigation and embodied question answering (EQA) persist due to inadequate modeling of fine-grained spatio-temporal cues and physical world comprehension. To address this, we propose VEME, a novel cross-modal alignment method that enhances generalization in unseen scenes by learning an ego-centric, experience-centered world model. Our framework integrates three key components: (1) a cross-modal alignment framework bridging objects, spatial representations, and visual semantics with spatio-temporal cues to enhance VLM in-context learning; (2) a dynamic, implicit cognitive map activated by world embedding to enable task-relevant geometric-semantic memory recall; and (3) an instruction-based navigation and reasoning framework leveraging embodied priors for long-term planning and efficient exploration. By embedding geometry-aware spatio-temporal episodic experiences, our method significantly improves reasoning and planning in dynamic environments. Experimental results on VSI-Bench and VLN-CE demonstrate 1%-3% accuracy and exploration efficiency improvement compared to traditional approaches.

cs.CV↗

arXiv2025

Contrastive Learning-Driven Traffic Sign Perception: Multi-Modal Fusion of Text and Vision

Traffic sign recognition, as a core component of autonomous driving perception systems, directly influences vehicle environmental awareness and driving safety. Current technologies face two significant challenges: first, the traffic sign dataset exhibits a pronounced long-tail distribution, resulting in a substantial decline in recognition performance of traditional convolutional networks when processing low-frequency and out-of-distribution classes; second, traffic signs in real-world scenarios are predominantly small targets with significant scale variations, making it difficult to extract multi-scale features.To overcome these issues, we propose a novel two-stage framework combining open-vocabulary detection and cross-modal learning. For traffic sign detection, our NanoVerse YOLO model integrates a reparameterizable vision-language path aggregation network (RepVL-PAN) and an SPD-Conv module to specifically enhance feature extraction for small, multi-scale targets. For traffic sign classification, we designed a Traffic Sign Recognition Multimodal Contrastive Learning model (TSR-MCL). By contrasting visual features from a Vision Transformer with semantic features from a rule-based BERT, TSR-MCL learns robust, frequency-independent representations, effectively mitigating class confusion caused by data imbalance. On the TT100K dataset, our method achieves a state-of-the-art 78.4% mAP in the long-tail detection task for all-class recognition. The model also obtains 91.8% accuracy and 88.9% recall, significantly outperforming mainstream algorithms and demonstrating superior accuracy and generalization in complex, open-world scenarios.

cs.CV↗

arXiv2015

Towards a solid solution of real-time fire and flame detection

Although the object detection and recognition has received growing attention for decades, a robust fire and flame detection method is rarely explored. This paper presents an empirical study, towards a general and solid approach to fast detect fire and flame in videos, with the applications in video surveillance and event retrieval. Our system consists of three cascaded steps: (1) candidate regions proposing by a background model, (2) fire region classifying with color-texture features and a dictionary of visual words, and (3) temporal verifying. The experimental evaluation and analysis are done for each step. We believe that it is a useful service to both academic research and real-world application. In addition, we release the software of the proposed system with the source code, as well as a public benchmark and data set, including 64 video clips covered both indoor and outdoor scenes under different conditions. We achieve an 82% Recall with 93% Precision on the data set, and greatly improve the performance by state-of-the-arts methods.

cs.CV↗