Institution profile

Tencent

Industry researchasia · cn
Official website
Research library1,252linked papers
Opportunities0open roles
Selected work

Representative Papers

A Survey of Deep Face Restoration: Denoise, Super-Resolution, Deblur, Artifact Removal

Nov 05, 2022arXiv.org

This paper presents a systematic survey of deep learning–based facial image restoration, focusing on denoising, super-resolution, deblurring, and artifact removal. Addressing challenges such as strong facial structural priors and complex degradation modeling, we propose the first holistic taxonomy of methods, a unified evaluation framework, and an open-source benchmark repository encompassing 20+ state-of-the-art approaches—including fully reproducible implementations. Leveraging datasets like CelebA and FFHQ, we conduct comprehensive cross-method evaluations using PSNR, SSIM, and LPIPS metrics, integrating CNN/Transformer architectures, perceptual and adversarial losses, and multi-scale feature fusion strategies. Our empirical analysis reveals performance boundaries and task-specific suitability across methods. Key contributions include: (1) the first structured, principle-driven classification system for facial restoration; (2) the first open, end-to-end benchmark platform supporting full method reproduction; (3) a rigorous, large-scale empirical study; and (4) concrete research directions concerning network design, evaluation paradigms, and dataset construction.

40 citations1 influentialRead paper

Decoupling Classifier for Boosting Few-shot Object Detection and Instance Segmentation

May 20, 2025Neural Information Processing Systems

This work addresses classification bias in few-shot object detection (FSOD) and few-shot instance segmentation (FSIS) caused by instance-level label omission—where unlabeled instances in images are erroneously treated as negative examples. To mitigate this, we propose a lightweight dual-head decoupled classifier that splits the standard classification head into two parallel branches: a positive branch for explicitly modeling ground-truth instances, and a noise-negative branch for modeling pseudo-negative instances arising from label omissions. Our approach is the first to formally define and address instance-level label-omission bias. It introduces no additional parameters, computational overhead, or reliance on meta-learning, fine-tuning, or external data; instead, it operates seamlessly within standard two-stage detectors (e.g., Mask R-CNN) under conventional supervised training. Extensive experiments on PASCAL VOC and MS-COCO benchmarks demonstrate consistent and significant improvements over state-of-the-art methods for both FSOD and FSIS. The code is publicly available.

26 citations3 influentialRead paper

Disambiguation of Chinese Polyphones in an End-to-End Framework with Semantic Features Extracted by Pre-Trained BERT

Sep 15, 2019Interspeech

This paper addresses the polyphonic character disambiguation challenge in Chinese end-to-end speech synthesis. We propose the first end-to-end grapheme-to-phoneme (G2P) framework for Chinese leveraging pretrained BERT: it takes raw character sequences as input—requiring neither manual word segmentation nor phoneme pre-annotation—and employs BERT to encode contextual semantics, jointly with FC, LSTM, and Transformer classifiers to predict polyphonic pronunciations. Our key contribution is the first application of BERT to Chinese polyphonic disambiguation, enabling semantic-driven, end-to-end modeling. Experiments demonstrate that BERT substantially improves disambiguation accuracy, outperforming an LSTM baseline on a standard test set. Furthermore, we empirically reveal that context window length critically influences disambiguation performance.

24 citationsRead paper

Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces

Jan 17, 2026

This work addresses the challenge that existing AI agent benchmarks inadequately evaluate performance on real-world, complex, and long-horizon command-line tasks. To bridge this gap, the authors introduce a novel evaluation benchmark comprising 89 high-difficulty terminal tasks, all derived from authentic workflows and accompanied by isolated execution environments, human-authored reference solutions, and automated verification tests. The benchmark is designed to ensure realism, verifiability, and diversity, substantially narrowing the disparity between practical scenarios and current model evaluation paradigms. Experimental results demonstrate that even state-of-the-art agents achieve success rates below 65% on this benchmark. The paper further provides comprehensive error analysis and publicly releases the dataset and evaluation toolchain to support future research in this domain.

9 citations1 influentialRead paper

AdaMARP: An Adaptive Multi-Agent Interaction Framework for General Immersive Role-Playing

Jan 16, 2026

This work addresses the limited immersion and dynamic adaptability of current large language models in role-playing scenarios, particularly their inability to support multi-agent coordination, scene transitions, and real-time character introduction. To overcome these limitations, the authors propose an adaptive multi-agent role-playing framework featuring a novel structured message format—[Thought]/(Action)/<Environment>/Speech—that integrates cognition, action, environment, and dialogue. An explicit scene manager is introduced to orchestrate narrative flow through discrete control actions and reasoning mechanisms. The study also contributes two specialized training datasets, AdaRPSet and AdaSMSet, along with AdaptiveBench, a trajectory-level evaluation benchmark. Experimental results demonstrate that an 8B-parameter role-playing model achieves superior character consistency and narrative coherence compared to several commercial large language models, while a 14B-parameter scene manager outperforms Claude Sonnet 4.5 in tasks involving dynamic character introduction and scene switching.

7 citationsRead paper
Recent publications

Latest Papers