🎯 Object Detection

🧪 ICML2026 · 7 paper notes

🔥 Top topics: Reasoning ×2 · Anomaly Detection ×2

HSGG: Training-Free Hierarchical Scene Graph Generation with Geometry-Guided Relation Reasoning

HSGG discovers entities from wholes to parts, propagates attributes from parts to wholes, and then combines geometric filtering with visually grounded predictions contrasted against geometry-induced hallucination priors, achieving 15.1 zR@100 on VG150 and 14.1 on PSG under SGDet without task-specific training (original Table 1).

OmniVerifier-M1: Multimodal Meta-Verifier with Explicit Structured Recalibration

Aiming at the issues that multimodal visual verifiers output binary signals (True/False) that are too coarse and that textual explanations are prone to reward-hacking, this paper proposes OmniVerifier-M1. It utilizes symbolic outputs such as bounding boxes as meta-verification rationales instead of text to support rule-based rewards like IoU. Theoretically and experimentally, it proves that decoupling binary judgment and meta-verification into two independent reward streams (rather than a multiplicative joint reward) significantly improves SNR. Ultimately, the verifier is upgraded to an agentic system, M1-TTS, capable of driving region-level self-recalibration.

EARL: Towards a Unified Analysis-Guided Reinforcement Learning Framework for Egocentric Interaction Reasoning and Pixel Grounding

EARL utilizes a two-stage MLLM framework of "coarse interpretation and fine response" to consolidate egocentric interaction reasoning tasks (description + Q&A + pixel mask) into a unified pipeline. The first stage outputs a global interaction description of the full image and treats the last hidden state as a semantic prior. This is injected into the second stage through a novel Analysis-guided Feature Synthesizer (AFS). The system is jointly trained via GRPO with a triple-reward mechanism (format/answer/grounding accuracy), outperforming Seg-Zero by 8.37% cIoU on Ego-IRGBench.

FOCUS: Forcing In-Context Object Localization through Visual Support Constraints and Policy Optimization

FOCUS uses a two-stage training approach—"complete removal of category names + attention mask optimization + GRPO IoU reward"—to force VLMs to perform in-context object localization based on visual support examples rather than semantic priors. The 7B parameter model outperforms 72B models, proving that task-aligned inductive bias is more important than pure scaling.

Testing the Test: Score-Direction Instability in Class-Split Anomaly Detection

The authors point out that "class-split" anomaly detection benchmarks are ill-posed when the anomaly class and the normal mixture distribution overlap in the representation space—AUROC collapses to random or even reverses, with the direction depending on the unknown anomaly class. A training-free "neighborhood class leakage" metric \(L_k\) is proposed to diagnose such benchmark failure before evaluation.

Adversarially Robust Approximate Furthest Neighbor

This theoretical paper provides the first approximate furthest neighbor data structure resistant to adaptive query adversaries. While maintaining a query complexity with \(n\)-dependence similar to Indyk's classical oblivious algorithm, it demonstrates that traditional random projection furthest neighbor algorithms can be broken by adaptive queries.

Mixture Prototype Flow Matching for Open-Set Supervised Anomaly Detection

MPFM replaces the traditional "unimodal Gaussian prototypes" in OSAD with a learnable Gaussian Mixture Model (GMM) prototype space. It uses flow matching to directly regress a velocity field in GMM form, augmented by a mutual information maximization regularization to prevent prototype collapse. The method outperforms all SOTA methods, including DRA, AHL, and DPDL, across 9 industrial and medical AD datasets under the 10/1 anomaly sample setting.