Tencent YouTu Lab had 14 papers accepted at ECCV 2024, covering diverse areas in computer vision including anomaly detection, image editing, and multimodal learning. Below are highlights of selected papers:

Multimodal Label Relevance Ranking via Reinforcement Learning

Proposes LR²PPO, a reinforcement learning approach for ranking label relevance by capturing human preferences. The method reduces annotation needs for new scenarios and introduces the LRMovieNet benchmark dataset.

DiffuMatting: Synthesizing Objects with Matting-level Annotation

A diffusion-based method that generates objects with precise alpha mattes using a green-screen inspired approach. The technique reduces MSE error by 15.4% for general objects and 11.4% for portraits.

OneNIP: Single Normal Image Prompt for Anomaly Detection

Introduces a reconstruction network using one normal image as a visual prompt to detect anomalies. Achieves state-of-the-art performance on MVTec, BTAD, and ViSA datasets.

Few-Shot Anomaly-Driven Generation for Industrial Inspection

Generates realistic anomalies from limited samples using a diffusion model. Improves AU-PR metrics by 5.8% for DRAEM and 1.5% for DesTSeg on MVTec.

Frequency Feature Fusion for Document Tamper Detection

Proposes FFDN with visual and wavelet-based frequency enhancement modules to detect subtle document forgeries. Outperforms existing methods on DocTamper dataset.

TF-FAS: Fine-Grained Semantic Guidance for Face Anti-Spoofing

Enhances generalization in face anti-spoofing through content and category element modules. Demonstrates superior cross-domain performance.

FaceAdapter: Fine-Grained Control for Diffusion Models

An efficient adapter for pretrained diffusion models enabling precise facial editing with identity preservation and attribute control.

Self-supervised Feature Adaptation for 3D Anomaly Detection

Introduces LSFA for adapting pretrained features to industrial 3D data, achieving 97.1% I-AUROC on MVTec-3D (+3.4% over prior work).

AdaCLIP: Hybrid Prompts for Zero-Shot Anomaly Detection

Combines static and dynamic prompts to adapt CLIP for zero-shot anomaly detection, showing strong generalization across 14 datasets.

RLR: Unified Reference Learning for Anomaly Detection

Prevents "learning shortcuts" in reconstruction-based methods through learnable reference representations. Outperforms alternatives on MVTec-AD and VisA.

Tuning-Free Image Customization with Multi-Modal Guidance

Enables precise region editing using both text and image prompts without fine-tuning. Achieves better results than methods requiring optimizaton.

TexDreamer: Zero-Shot 3D Human Texture Generation

Generates high-fidelity 3D human textures from text or images in seconds, accompanied by the ATLAS dataset of 50K textures.

AccDiffusion: High-Resolution Image Generation

Solves object repetition in patch-based high-res generation through content-aware prompts and dilated sampling with window attention.

FreeMotion: Unified Framework for Multi-Person Motion Synthesis

Generalizes text-to-motion generation to arbitrary numbers of people through conditional motion distribution modeling.