Highlights from CVPR 2026: Four NIRC Papers on Video Anomaly Detection and Hand Modeling

CVPR 2026 Overview

From June 3‑7, 2026, the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) was held in Denver, Colorado. Recognized by the China Computer Federation as a CCF A‑class venue, the conference recorded unprecedented numbers of authors, reviewers, and area chairs, and highlighted generative models, 3‑D vision, video understanding, and embodied intelligence as key topics.

Four NIRC Papers Presented

1. Fine‑VAD: Fine‑Grained Video Anomaly Detection

The paper introduces a progressive cross‑granularity learning framework that moves from binary anomaly labels to pseudo‑macro categories and finally to fine‑grained class labels, aligning video features with increasingly detailed semantic supervision. Experiments on multiple benchmarks show significant gains over prior methods and demonstrate that the approach is not tied to a single model architecture.

2. Alert‑CLIP

Alert‑CLIP proposes an abnormality‑aware tuning of CLIP for video anomaly detection. It addresses CLIP’s difficulty in distinguishing normal from abnormal semantics by performing three‑level alignment: video‑label, region‑text, and region‑semantic. The authors also release the VAGTA dataset, containing 4,212 video clips with region annotations and textual descriptions. Results indicate superior performance in weak‑supervised, zero‑shot, and open‑vocabulary scenarios.

3. Clay‑to‑Stone: Phase‑Wise 3D Gaussian Splatting for Monocular Hand‑Object Manipulation

This work presents a two‑stage dynamic articulated modeling pipeline. First, a semantically constrained free‑form deformation explores latent structure and motion patterns; second, rigid physical constraints solidify the structure and explicitly estimate hinge parameters, achieving physically consistent high‑quality reconstructions of articulated objects from monocular input.

4. Temporal and Content Co‑Awareness Latent Diffusion for Controllable Hand Image Generation

The authors identify that pose and appearance modulation in diffusion models depend jointly on the denoising timestep and the complexity of conditioning signals. They design a dynamic modulation mechanism that adaptively allocates injection strength for pose and appearance cues, resulting in improved pose consistency and visual fidelity compared with existing methods.

Poster Sessions and Workshops

During the poster exhibition, the author presented the controllable hand image generation work, engaging in concise explanations and receiving valuable feedback. The pre‑conference workshops offered a focused view of frontier directions, discussing unresolved challenges and future research pathways beyond the main conference papers.

Networking and Community Interaction

Face‑to‑face conversations with authors of previously known works allowed deeper insight into research motivations, design trade‑offs, experimental details, and prospective directions—information often omitted from papers. Informal exchanges at poster areas and between sessions sparked many spontaneous ideas.

Experiences in Denver

The opening ceremony highlighted the rapid growth of the computer‑vision field. Outside the conference, the author explored downtown Denver, noting its relaxed atmosphere, wide streets, abundant sunshine, and striking evening skies. Evening events included a lively music dinner and a memorable “kneeling‑talk” poster interaction.