Research & Projects
This page records my research training and engineering projects. For team research, it describes only my confirmed responsibilities in experimental implementation, evaluation, and analysis.
Research training
Experimental Study on Multi-turn Multimodal Fashion Image Retrieval
- Contributing to experimental validation and performance analysis for multi-turn interactive image retrieval in realistic shopping scenarios, covering text-to-image, composed image retrieval, in-shop retrieval, and sketch-to-image tasks.
- Participating in the FashionAM three-stage training and evaluation pipeline, including background-removed image alignment, fashion-caption alignment, MLLM + LoRA multi-turn query modeling, and multi-positive contrastive learning.
- Analyzing R@K and mAP on DIM-Fashion and MT FashionIQ settings, with attention to multi-task, multi-turn-context, and intent-switching behavior.
- Reproducing experiments and ablation studies to analyze the effects of training stages, alignment objectives, and MLLM backbones on retrieval performance.
Experimental Study on Diffusion Model Acceleration and Artifact Analysis
- Participated in a RALU-based diffusion-model acceleration study on grid-like artifacts caused by staged latent upsampling in high-frequency image regions.
- Implemented and evaluated FLUX.1-dev experiments with a Stage-3 VAE re-projection comparison pipeline, including VAE decoding, pixel-space upsampling, VAE encoding, patch-coordinate mapping, and re-noising settings.
- Compared FID, CLIP, GenEval, CLIP-IQA, and inference latency on settings including cc12m and COCO2017 to examine quality-efficiency trade-offs.
- Observed that VAE re-projection can reduce high-frequency grid artifacts, but introduces extra encoding-decoding cost and reduced sharpness, so it does not directly preserve the original acceleration advantage.
Robotics vision and engineering projects
RoboMaster Robotics Competition / ARTINX Lab
- Led model training, inference deployment, and on-robot debugging for the robot vision system across hero, infantry, and sentry platforms.
- Fine-tuned YOLO-series object detectors with PyTorch and deployed them to Jetson Orin NX via TensorRT for armor-plate detection and localization. Data augmentation and model replacement improved overall accuracy by about 15% below 20 Lux while maintaining typical latency within 10 ms.
- Designed a periodic observation and timed-shooting algorithm for the outpost target, supporting stable hits on dynamic targets with average hit efficiency of about 90%.
- Used Extended Kalman Filter-based target-motion observation and control-aware aiming-direction selection. Also contributed to a C++ Actor Framework visualization module and ArtinxHub maintenance.
VLM Image Captioning and Component Analysis
- Implemented an end-to-end LLaVA-style vision-language model with PyTorch and Transformers, integrating ViT/ResNet visual encoders with GPT-2/Qwen language models for image-captioning training and evaluation.
- Built the training, inference, and evaluation pipeline on COCO, using BLEU and CIDEr for quantitative analysis.
- Conducted ablations on the visual encoder, connector, and language model components. An attention connector improved CIDEr by more than 230% over an MLP connector in the experiment.
Face Recognition and Expression Interaction System on Jetson Nano
- Built a real-time face-detection, face-recognition, and expression-interaction system on a Jetson Nano 4GB board, achieving approximately 30 fps end-to-end.
- Used RetinaFace-500MF with a MobileNet backbone for detection, MobileFaceNet@WebFace600K for recognition, and MobileNetV3 fine-tuned on FER2013 for expression recognition.
- Applied TensorRT and Jetson ecosystem tools for edge inference optimization under limited compute and memory resources.