CV
Education
- M.S. in Information and Communication Engineering, Dalian University of Technology, 2024.09—2027.07
- Research focus: end-to-end autonomous driving, world models, computer vision, diffusion models, and multimodal AI
- B.S. in Artificial Intelligence, Dalian University of Technology, 2020.09—2024.07
- Major courses: Computer Vision, Deep Learning, Linear Algebra, and Image Processing
- Recommended for postgraduate study to the Lu Huchuan research group
Experience
- 2026.06—2026.09: Algorithm Intern, Bosch (China) Investment Co., Ltd. — Intelligent Driving & Control (XC)
- Worked on algorithm design, optimization, evaluation, and real-vehicle deployment for one-stage end-to-end world models.
- Built an internal world-model foundation on FastWAM and designed DeltaWorld for dynamic-environment understanding, reducing trajectory prediction error by 30%.
- Developed PV-Register to improve traffic-light representation and interaction; increased interaction success from 20% to 90%.
- Led temporal-modeling iterations that reduced real-vehicle trajectory-jitter cases by 90% and improved acceleration/braking comfort.
- Designed a single-step trajectory-generation inference strategy, reducing latency from 85 ms to 20 ms; contributed to automated evaluation and on-vehicle debugging.
- 2025.09—2026.05: Model Intern, Qingzhou Zhihang Technology Co., Ltd. — Parking Model Team
- Took primary ownership of building a VAD end-to-end scoring model from scratch, including Encoder—Decoder architecture, task-specific encoders for detection/OCC/navigation, and data augmentation.
- Applied Focal Loss + OHEM, R-Drop, and self-supervised learning to address long-tail and overfitting issues; rare/complex-scene trajectory pass rate improved from 30% to 85%.
- Designed a trajectory difficulty vocabulary and reusable evaluation metrics for team-wide iteration.
- Designed collision losses and a classification RL strategy, reducing scenario collision rate from 2% to 0.3%; introduced Recurrent temporal modeling, reducing trajectory-consistency P95 from 16 m to 3 m.
- Processed data at the scale of tens of millions of frames and completed exploratory studies of Diffusion Planner and MotionLM.
Research & Projects
- DreamMix: Decoupling Object Attributes for Enhanced Editability in Customized Image Inpainting — IJCV (CCF-A), First Author
- Proposed an attribute-decoupling mechanism, textual attribute substitution module, and local-global disentangled inpainting framework to improve identity preservation and attribute editability.
- Achieved CLIP-T 0.289 on SDXL and 0.290 on FLUX, outperforming representative baselines; adapts with only 4—6 reference images.
- Open-source project with 100+ GitHub stars. See the publication page.
- FineRS++: Multi-Hypothesis Reasoning and Segmentation of Small Objects in High-Resolution Images — TPAMI, Under Review, Student First Author
- Built a Qwen2.5-VL-based global-local fine-grained perception framework with coarse-to-fine MLLM reinforcement learning (GSE + LPR) and a GRPO localization-aware backtracking reward.
- Constructed a dedicated high-resolution small-object annotation dataset and achieved gIoU 55.1, cIoU 46.5, MVQA 83.3%, and OVQA 56.7% on FINERS-4k.
- Pet Style Transformation — FLUX, Image Translation, Image Customization, LoRA; Project Lead
- Built a Flux-Fill in-context generation and paired-data synthesis pipeline with LoRA weight fusion, enabling pet-preserving cute-style transfer with cross-sample generalization.
- MoE-based Object Detection for Challenging Environments — MoE, DINO, Object Detection; Project Lead
- Added a global-local routing MoE module to DINO and built the dataset pipeline; improved mAP from 0.52 to 0.59 in night, rain, snow, and other challenging conditions.
Awards
- RoboMaster Super Competition — National First Prize, Vision Team Deputy Leader (auto-aim, object detection, 3D SLAM)
- China Robot and Artificial Intelligence Competition — National First Prize, Team Captain (2D SLAM, robotic-arm control and planning)
- RoboCup China Robot Competition — National First Prize, Team Captain (instance segmentation, edge deployment, geometric planning)
- National College Student Innovation and Entrepreneurship Program — Provincial First Prize, Team Captain (voice interaction, pose recognition, IoT)
- 10+ national and provincial awards, including a National Second Prize in the China College Computer Design Competition and the China University Intelligent Robot Creative Competition
Technical Skills
- Deep learning: PyTorch, diffusion models, multimodal LLMs, reinforcement learning, model fine-tuning, LoRA
- Computer vision: image generation/editing, object detection, segmentation, BEV, visual reasoning, temporal modeling
- Robotics & deployment: 2D/3D SLAM, auto-aim, robotic-arm planning, edge/vehicle deployment, automated evaluation
- Programming: Python, C/C++, Linux, OpenCV
Contact
- Email: 2286247133@qq.com
- Location: Dalian, China
- Phone: available on request
Publications
FineRS++: Multi-Hypothesis Reasoning and Segmentation of Small Objects in High-Resolution Images
YiCheng Yang et al. “FineRS++: Multi-Hypothesis Reasoning and Segmentation of Small Objects in High-Resolution Images.” Under review at TPAMI.
DreamMix: Decoupling Object Attributes for Enhanced Editability in Customized Image Inpainting Permalink
YiCheng Yang et al. “DreamMix: Decoupling Object Attributes for Enhanced Editability in Customized Image Inpainting.” International Journal of Computer Vision.