BruceFeng98/AiDLab

--- license: apache-2.0 tags: - fashion-ai - computer-vision - generative-ai size_categories: - 1K<n<10K viewer: false --- <img src="https://huggingface.co/datasets/BruceFeng98/AiDLab/resolve/main/AiDLab.png" width="1400" /></img> # Recruitment: Research Assistant (RA) @ AiDLab, Hong Kong 香港AiDLab 招聘研究助理 **Contact**: `[e-mail]` --- ## 人工智能设计研究所 (AiDLab) - **负责人**: AiDLab 总裁黄伟强教授 (Prof. Wai Keung Wong)，香港理工大学郑翼雄时装教授 - **地点**: 香港科学园 (Hong Kong Science Park) --- ## 📋 工作内容与要求 - 开展计算机视觉 (CV) 前沿算法研究及落地场景探索 - 调研特定场景的算法落地可行性，负责算法的深度研究与优化 - 推动科研成果在时尚设计领域的实际应用 1. **学历背景**: 图像处理、计算机视觉、自然语言处理等相关方向，本科及以上学历 2. **技术实力**: - 扎实的 Python / 代码基础，具备快速复现前沿论文算法的能力 - 熟悉生成式算法 (Generative AI) 的优缺点，对生成式任务有浓厚兴趣 3. **软实力**: 对 Fashion（时尚）领域感兴趣，具备良好的审美或行业洞察力 --- ## 🌟 Why Join Us - **地理位置优越**: 位于香港科学园，配套设施完善 - **产研结合**: 实验室已有成熟的落地商用产品 - **学术氛围**: 与理大及 AiDLab 的专家共同工作 --- ## 🎥 FashionShow实时视频生成课题（可运程） 可选合作开展形式 - （1）香港线下RA+可支持读博 - （2）远程科研合作+实习津贴+可共一署名 核心工作内容 - （1）视频数据处理体系和视频生成测评体系 - （2）高效视频自回归生成模型 <video src="https://huggingface.co/datasets/BruceFeng98/AiDLab/resolve/main/runwayshots.mp4" width="1024" controls autoplay loop/></video> --- ## 🧠 技术命题与深度讨论 (Research Challenges) 我们诚邀对以下 Fashion AI 前沿问题有独到见解的同学加入讨论，选择一两个问题，你可以先让AI思考，然后加入自己的见解，也可提出自己其他的问题思考： ### 1. 伪高清视频的识别与视觉质量评估 * **挑战**：许多历史影像虽经插值提升了分辨率（如 4K），但实际视觉质量（清晰度、纹理细节）依然很差。 * **思考**：如何构建 **无参考视频质量评估 (No-Reference VQA)** 模型？ ### 2. 视频中运动模糊 (Motion Blur) 的解决路径 * **挑战**：T 台走秀中模特动作较快，采集到的数据常伴随严重的运动模糊。 * **思考**：在生成式模型中，是应该在**预处理阶段**引入去模糊算法（Deblurring），还是在 **Diffusion Model 的训练阶段** 引入运动轨迹先验（Motion Prior）来增强对模糊帧的重建能力？ ### 3. FashionShow 复杂镜头语言的受控生成 * **挑战**：秀场包含推拉摇移、侧拍、俯拍等极其丰富的摄影机轨迹。 * **思考**：如何解耦**人体运动**与**相机运动**？是否可以通过引入 Camera Injection（如 CameraCtrl）或参考位姿序列（Pose Sequence）来实现对特定镜头语言的精准复现？ ### 4. 复杂背景下的人像前后景分离 * **挑战**：T 台周围常有密集的观众和相似的模特背景，传统分割易出现粘连。 * **思考**：在 Fashion 场景下，如何结合 **Robust Video Matting (RVM)** 与最新的 **Segment Anything (SAM 2)** 提升时序分割的稳定性？针对前景后景都是人像的极端情况，如何利用深度估计（Depth Estimation）进行语义层级的遮挡关系建模？ --- ## 📩 投递通道 **投递邮箱**：[[e-mail]](mailto:[e-mail]) --- `#ComputerVision` `#AIGC` `#FashionAI` `#HongKongJobs` `#AiDLab`license: apache-2.0 tags: - fashion-ai - computer-vision - generative-ai size_categories: - 1K<n<10K viewer: false --- <img src="https://huggingface.co/datasets/BruceFeng98/AiDLab/resolve/main/AiDLab.png" width="1400" /></img> # Recruitment: Research Assistant (RA) @ AiDLab, Hong Kong **Contact**: `[e-mail]` --- ## AI Design Institute (AiDLab) - **Head**: Prof. Wai Keung Wong, President of AiDLab, Cheng Yik Hung Professor in Fashion, The Hong Kong Polytechnic University - **Location**: Hong Kong Science Park --- ## 📋 Job Responsibilities and Requirements - Conduct cutting-edge computer vision (CV) algorithm research and explore practical deployment scenarios - Investigate the feasibility of algorithm deployment in specific scenarios, conduct in-depth research and optimization of algorithms - Promote the practical application of research findings in the fashion design field 1. **Educational Background**: Bachelor's degree or above in fields related to image processing, computer vision, natural language processing, etc. 2. **Technical Competencies**: - Solid Python/programming foundation, with the ability to quickly reproduce cutting-edge paper algorithms - Familiar with the pros and cons of generative AI, with strong interest in generative tasks 3. **Soft Skills**: Interested in the fashion field, with good aesthetic taste or industry insights --- ## 🌟 Why Join Our Team - **Prime Location**: Located at Hong Kong Science Park with complete supporting facilities - **Industry-Academia Integration**: The lab has mature commercialized products - **Academic Atmosphere**: Collaborate with experts from The Hong Kong Polytechnic University and AiDLab --- ## 🎥 FashionShow Real-time Video Generation Project (Remote Work Available) Optional Cooperation Forms - (1) Full-time RA in Hong Kong + PhD application support - (2) Remote research cooperation + internship stipend + co-first authorship opportunity Core Job Responsibilities - (1) Video data processing system and video generation evaluation system - (2) Efficient video autoregressive generation models <video src="https://huggingface.co/datasets/BruceFeng98/AiDLab/resolve/main/runwayshots.mp4" width="1024" controls autoplay loop/></video> --- ## 🧠 Research Challenges and In-depth Discussions We sincerely invite students with unique insights into the following cutting-edge Fashion AI issues to join the discussions. You can select one or two topics: first let AI think, then add your own insights, or raise other questions for consideration: ### 1. Pseudo-High-Definition Video Recognition and Visual Quality Assessment * **Challenge**: Many historical videos have been upscaled (e.g., to 4K resolution) via interpolation, but their actual visual quality (clarity, texture details) remains poor. * **Thought**: How to build a No-Reference Video Quality Assessment (No-Reference VQA) model? ### 2. Solutions for Motion Blur in Videos * **Challenge**: Models walk fast on fashion runways, leading to severe motion blur in captured data. * **Thought**: In generative models, should we introduce deblurring algorithms in the preprocessing stage, or introduce motion priors during the training phase of Diffusion Models to enhance the ability to reconstruct blurred frames? ### 3. Controlled Generation of Complex Cinematic Language in Fashion Shows * **Challenge**: Fashion shows feature extremely diverse camera movements, including pushing, pulling, panning, tilting, side shooting, overhead shooting, etc. * **Thought**: How to decouple human motion and camera motion? Can we achieve accurate reproduction of specific cinematic language by introducing Camera Injection (e.g., CameraCtrl) or reference pose sequences? ### 4. Foreground-Background Separation of Portraits in Complex Backgrounds * **Challenge**: There are often dense audiences and similar model backgrounds around fashion runways, leading to adhesion issues in traditional segmentation. * **Thought**: In the fashion scenario, how to combine Robust Video Matting (RVM) and the latest Segment Anything (SAM 2) to improve the stability of temporal segmentation? For the extreme case where both foreground and background are portraits, how to use depth estimation to model occlusion relationships at the semantic level? --- ## 📩 Application Channel **Application Email**: [[e-mail]](mailto:[e-mail]) --- `#ComputerVision` `#AIGC` `#FashionAI` `#HongKongJobs` `#AiDLab`