Xiaowei Song   宋枭炜

I am a Second-year Computer Science Master's student at CAD Research Center, Tongji University.

I did my B.Sc. in Compter Science and Technology at Shanghai University.

My research interests are computer vision, machine learning and multimodal generation, specifically, learning-based methods for 3D implicit representation, 3D Gaussian Splatting, Generation, etc.

If you find any research interests that we might share, feel free to drop me an email. I am always open to potential collaborations.

Email  /  CV  /  Github /  Linkedin

中文 / English

Much of my research is about inferring the physical world (shape, motion, color, light, etc) from images and 3D raw data. Representative papers are highlighted.

A training-free approach that can be directly applied to the inference process of any pretrained 3DGS model to resolve its visual artefacts at drastically changed rendering settings.

Using AutoRF, I replicated its performance on the KITTI360 dataset. This involved data type loading, camera coordinate system conversion and NOCS spatial transformation.

We gather the wiki pedia knowledge about science questions to make it into a OpenBook Q&A question, then we make three Deberta models with different finetuning and combine their output features to infer the right answer.

We use a integration of ViT OFA and BLIP to make predictions on a dataset containing a wide variety of (prompt, image) pairs generated by Stable Diffusion 2.0.

This project is about HUAWEI robots application, the project requires us to assign policies, control scheduling, and path planning for multiple robots in a single map.

To study the problem of inconsistency between the regions of interest of the classification head and localization head for single-stage target detection. Propose (1) a training strategy for positive sample point regression frame center clustering and positive sample point semantic center weighting (2) a consistent optimization method for quadratic regression optimization frame quality. Eventually the proposed method outperformed the baseline in both accuracy and recall and became the SOTA for the task at that time.