🎯 Object Detection

🤖 AAAI2026 · 29 paper notes

📌 Same area in other venues: 📷 CVPR2026 (97) · 🔬 ICLR2026 (31) · 🧪 ICML2026 (6) · 🧠 NeurIPS2025 (27) · 📹 ICCV2025 (28)

🔥 Top topics: Anomaly Detection ×9 · Object Detection ×8 · Remote Sensing ×3 · Few-/Zero-Shot Learning ×3 · Object Tracking ×2

AerialMind: Towards Referring Multi-Object Tracking in UAV Scenarios

This paper introduces AerialMind, the first large-scale Referring Multi-Object Tracking (RMOT) benchmark dataset for UAV scenarios, and proposes HawkEyeTrack (HETrack), a method that achieves language-guided multi-object tracking in aerial UAV scenes via a co-evolutionary fusion encoder and a scale-adaptive contextual refinement module.

An Overall Real-Time Mechanism for Classification and Quality Evaluation of Rice

This paper proposes a real-time overall mechanism for rice quality evaluation, integrating three modules: an improved YOLO-v5 (variety detection), an improved ConvNeXt-Tiny (intactness grading), and K-means (chalkiness region quantification). The system achieves 99.14% mAP and 97.89% detection accuracy on a self-constructed dataset of 20,000 images spanning six rice varieties.

AnoStyler: Text-Driven Localized Anomaly Generation via Lightweight Style Transfer

This work formulates zero-shot anomaly generation as a text-guided localized style transfer problem. A lightweight U-Net trained with CLIP-based losses stylizes masked regions of normal images into semantically aligned anomalous images. With only 263M total parameters (0.61M trainable), AnoStyler surpasses diffusion-based baselines on MVTec-AD and VisA while significantly improving downstream anomaly detection performance.

AquaSentinel: Next-Generation AI System Integrating Sensor Networks for Urban Underground Water Pipeline Anomaly Detection via Collaborative MoE-LLM Agent Architecture

This paper proposes AquaSentinel, a physics-informed AI system that achieves network-wide pipeline leak detection using only 20–30% node coverage through sparse sensor deployment, physics-augmented virtual sensors, a MoE spatiotemporal GNN ensemble, a dual-threshold RTCA detection algorithm, causal flow localization, and LLM-based report generation. The system achieves 100% detection rate across 110 leak scenarios.

Beyond Boundaries: Leveraging Vision Foundation Models for Source-Free Object Detection

This paper proposes a framework leveraging VFMs (DINOv2 + Grounding DINO) to enhance Source-Free Object Detection (SFOD) via three modules: Patch-weighted Global Feature Alignment (PGFA), Prototype-based Instance Feature Alignment (PIFA), and Dual-source Enhanced Pseudo-label Fusion (DEPF). The method achieves state-of-the-art results on 6 cross-domain detection benchmarks, e.g., 47.1% mAP on Cityscapes→Foggy Cityscapes (+3.5% over DRU) and 67.4% AP on Sim10k→Cityscapes (+8.7% over DRU).

CASL: Curvature-Augmented Self-supervised Learning for 3D Anomaly Detection

This work identifies point cloud curvature as a powerful cue for anomaly detection and proposes CASL, a curvature-augmented self-supervised learning framework. By guiding coordinate reconstruction with multi-scale curvature prompts, CASL learns generalizable 3D representations without any anomaly-detection-specific mechanisms, achieving a 5.6% O-AUROC improvement over the previous state of the art on Real3D-AD.

Commonality in Few: Few-Shot Multimodal Anomaly Detection via Hypergraph-Enhanced Memory

This paper proposes CIF, which leverages hypergraphs to extract intra-class structural commonalities from a small number of training samples, guiding memory bank construction and retrieval for few-shot multimodal industrial anomaly detection, achieving state-of-the-art performance.

Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning

This paper proposes GroundingAgent, a visual grounding framework that requires no task-specific fine-tuning. By composing pretrained open-vocabulary detectors (YOLO World), an MLLM (Llama-3.2-11B-Vision), and an LLM (DeepSeek-V3) into a structured iterative reasoning pipeline, the method achieves a zero-shot average accuracy of 65.1% on RefCOCO/+/g, substantially outperforming prior zero-shot approaches.

Correcting False Alarms from Unseen: Adapting Graph Anomaly Detectors at Test Time

This paper proposes TUNE, a plug-and-play test-time adaptation framework that addresses the "normality shift" problem in graph anomaly detection—caused by the emergence of new normal node categories—by transforming node features via a graph aligner. It leverages the degree of aggregation contamination as an unsupervised adaptation signal and significantly enhances the generalization of various pretrained GAD models across 10 real-world datasets.

CountSteer: Steering Attention for Object Counting in Diffusion Models

This paper proposes CountSteer, a training-free inference-time method that injects adaptive steering vectors into the cross-attention hidden states of diffusion models, improving object counting accuracy by approximately 4% without degrading image quality.

FDP: A Frequency-Decomposition Preprocessing Pipeline for Unsupervised Anomaly Detection in Brain MRI

This work presents the first systematic frequency-domain analysis of brain MRI anomalies, demonstrating that lesions are predominantly concentrated in low-frequency components. Based on this finding, the authors propose the Frequency Decomposition Preprocessing (FDP) framework, which reconstructs low-frequency signals via a learnable prior context bank to suppress lesions while preserving anatomical structures. As a plug-and-play module, FDP consistently improves detection performance across multiple UAD baselines (achieving a 17.63% DICE gain on LDM).

Harnessing Vision-Language Models for Time Series Anomaly Detection

A two-stage zero-shot time series anomaly detection framework is proposed: ViT4TS employs a lightweight ViT to perform multi-scale cross-patch matching on line-chart renderings of time series for candidate anomaly interval localization, while VLM4TS leverages GPT-4o with global temporal context to validate and refine detection results. The framework achieves F1-max surpassing the best baseline by 24.6% across 11 benchmarks, with token consumption only 1/36 of existing LLM-based methods.

LampQ: Towards Accurate Layer-wise Mixed Precision Quantization for Vision Transformers

This paper proposes LampQ, a metric-based layer-wise mixed precision quantization method that measures the quantization sensitivity of each ViT layer via a type-aware Fisher information metric, combines integer linear programming to optimize bit-width allocation, and iteratively refines the allocation. LampQ achieves state-of-the-art performance across image classification, object detection, and zero-shot quantization tasks.

LoReTTA: A Low Resource Framework To Poison Continuous Time Dynamic Graphs

This paper proposes LoReTTA, a two-stage adversarial poisoning attack framework that requires no surrogate model. It first sparsifies high-influence edges via 16 temporal importance metrics, then replaces them with adversarial edges using a degree-preserving negative sampling algorithm. Across 4 datasets × 4 TGNN models, LoReTTA achieves an average performance degradation of 29.47%, while evading 4 anomaly detection systems and resisting 4 defense methods.

MovSemCL: Movement-Semantics Contrastive Learning for Trajectory Similarity (Extension)

This paper proposes MovSemCL, a framework that transforms GPS trajectories into movement-semantic features (displacement vectors + heading angles + Node2Vec spatial graph embeddings), achieves hierarchical encoding via patch-level two-stage attention (reducing complexity from \(O(L^2)\) to near-linear), and designs Curvature-Guided Augmentation (CGA) to preserve behaviorally critical segments such as turns and intersections. The framework achieves a mean rank approaching the ideal value of 1 on trajectory retrieval tasks while reducing inference latency by 43.4%.

CountVid: Open-World Object Counting in Videos

This paper proposes CountVid, a model, and the VideoCount benchmark, presenting the first systematic study of open-world video object counting—given a text or image description specifying target objects, the system enumerates all unique instances in a video. By combining an image counting model with a promptable video segmentation and tracking model, CountVid addresses challenges such as occlusion and re-appearance, achieving substantial improvements over strong baselines across diverse scenarios including TAO, MOT20, penguin colonies, and X-ray metal crystallization.

PromptMoE: Generalizable Zero-Shot Anomaly Detection via Visually-Guided Prompt Mixing of Experts

PromptMoE shifts prompt learning from a monolithic paradigm to a compositional one. Through a visually-guided Mixture of Experts (MoE) mechanism, it dynamically assembles instance-adaptive normal/abnormal state prompts from a learnable semantic primitive bank, achieving state-of-the-art zero-shot anomaly detection (ZSAD) performance across 15 industrial and medical datasets.

RcAE: Recursive Reconstruction Framework for Unsupervised Industrial Anomaly Detection

This paper proposes a Recursive Convolutional Autoencoder (RcAE) that progressively suppresses anomalies while preserving normal details through multi-step iterative reconstruction with shared parameters. Combined with a Cross-Recursive Detection module (CRD) that exploits multi-step reconstruction dynamics for robust anomaly localization, the method achieves performance comparable to state-of-the-art approaches using only 10% of the parameters required by diffusion models.

Reimagining Anomalies: What if Anomalies Were Normal?

This paper proposes the first counterfactual explanation framework for unsupervised image anomaly detection. By training a generator to modify anomalous samples into multiple disentangled counterfactuals perceived as normal by the detector, the framework answers at the semantic level: "What would an anomaly look like if it were normal?" This provides a depth of interpretability far exceeding traditional heatmap-based approaches.

REXO: Indoor Multi-View Radar Object Detection via 3D Bounding Box Diffusion

This paper extends the 2D bounding box diffusion paradigm of DiffusionDet to 3D radar space, proposing the REXO framework. It enables explicit cross-view radar feature association guided by noisy 3D bounding box projections, and introduces a ground-level constraint to reduce the diffusion parameter space. REXO surpasses the state of the art by +4.22 AP and +11.02 AP on the HIBER and MMVR indoor radar datasets, respectively.

SimROD: A Simple Baseline for Raw Object Detection with Global and Local Enhancements

This paper proposes SimROD, an extremely lightweight (only 0.003M parameters) RAW image object detection method that surpasses complex state-of-the-art approaches on multiple RAW detection benchmarks through global Gamma enhancement (4 learnable parameters) and green channel-guided local enhancement.

SM3Det: A Unified Model for Multi-Modal Remote Sensing Object Detection

SM3Det introduces the M2Det task for remote sensing (multi-modal datasets + multi-task object detection), employing a grid-level sparse MoE backbone and a Dynamic Sub-module Optimization (DSO) mechanism to handle SAR/optical/infrared modalities with both horizontal and oriented bounding box detection in a single unified model, substantially outperforming three independently trained modality-specific models combined.

T-Rex-Omni: Integrating Negative Visual Prompt in Generic Object Detection

This paper proposes T-Rex-Omni, the first framework to systematically incorporate negative visual prompts into open-set object detection. Through a training-free NNC module and an NNH loss, it substantially narrows the performance gap between visual-prompt and text-prompt detection methods, with particularly strong results in long-tail scenarios (LVIS-minival APr of 51.2).

Temporal Object-Aware Vision Transformer for Few-Shot Video Object Detection

This paper proposes an object-aware temporal modeling framework that achieves cross-frame temporal consistency through selective propagation of high-confidence detection features. Combined with a pretrained vision-language encoder (OWL-ViT) and a few-shot detection head, the method achieves an average improvement of 3.7%–5.3% AP across four video few-shot detection benchmarks.

Towards Multiple Missing Values-Resistant Unsupervised Graph Anomaly Detection

This paper proposes M2V-UGAD, the first framework to address unsupervised graph anomaly detection under simultaneous node attribute and graph topology missingness. Through three core mechanisms—dual-pathway independent imputation, hyperspherical latent space fusion, and pseudo-anomaly generation—the framework overcomes cross-view interference and imputation bias, consistently outperforming existing methods across 7 benchmark datasets.

TubeRMC: Tube-conditioned Reconstruction with Mutual Constraints for Weakly-supervised Spatio-Temporal Video Grounding

This paper proposes TubeRMC, a framework that generates text-conditioned candidate tubes and performs tube-conditioned reconstruction along temporal, spatial, and spatio-temporal dimensions, augmented by spatial-temporal mutual constraints to improve weakly-supervised spatio-temporal video grounding.

VK-Det: Visual Knowledge Guided Prototype Learning for Open-Vocabulary Aerial Object Detection

VK-Det is proposed as a framework that leverages only the visual knowledge of VLMs (without any additional supervision signals) to achieve state-of-the-art performance in open-vocabulary aerial object detection through Adaptive Selection Knowledge Distillation (ASKD), Prototype-Aware Pseudo-Label generation (PAPL), and Synthetic Matching Inference (SMI), even surpassing methods that rely on extra supervision.

When Trackers Date Fish: A Benchmark and Framework for Underwater Multiple Fish Tracking

This paper proposes MFT25, a large-scale underwater multiple fish tracking dataset (15 sequences, 408K annotations), and SU-T, a tracking framework combining UKF with FishIoU, achieving state-of-the-art performance of 34.1 HOTA and 44.6 IDF1. Statistical analyses further reveal fundamental differences between fish tracking and terrestrial object tracking.

YOLO-IOD: Towards Real Time Incremental Object Detection

This work is the first to systematically integrate incremental object detection (IOD) into the YOLO real-time framework. It identifies three types of knowledge conflict, proposes a three-module solution (CPR + IKS + CAKD), and introduces the more realistic LoCo COCO benchmark for evaluation.