Xiaohan Yan   颜小涵

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

I did my B.Sc. in Computing Science at Hohai University.

My research interests are computer vision, multi modal and reinforcement learning, specifically, learning-based methods for 3D point cloud segmentation, multimodal pretrain methods, 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.

I am a former ACMer and a former OIer.

Email  /  CV  /  Github /  Linkedin 

中文 / English / 日本語

More details has been written in my CV.

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

We leverage the 3D geometry information in 3D point cloud, the projection relationship between 3D point cloud and multi-view 2D posed RGB-D frames and the semantic features extracted by CLIP from multi-view 2D posed RGB-D frames to address the challenge of 3D instance segmentation.

Similar to how humans perceive 3D objects, neural networks discern the class labels of point clouds by combining local and global features of the structures and performance. Based on this, we reviewed the pipeline of few-shot point cloud semantic segmentation and identified three issues.

Meta-learning plays an increasingly importantt role in AutoML. A key sub-problem—meta-learning from learning curves is an mmature but gradually attention area within the field of meta-learning.

Learning medical visual representations through vision-language pre-training has some challenges, i.e., local alignment lacks interpretability and clinical relevance, and the insufficient internal and external representation learning of image-report pairs. To address these issues, we propose an Anatomical Structure-Guided (ASG) framework.

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

This project aims to using Segment Anything 3D to solve a ply pointcloud without 2D labels.

We using a pcd2rgb method to generate 2D RGB & Depth photos. And then, we aligne the inputs and generate the .ply output.

We get the 0.905 at the leardboard. And reach the Top 4%.

We gather the wiki pedia knowledge about science questions, and use the bag-of-words model to clean the datas. Then, we use the sentence transformer to find the similarity between the problen and the cleaned dataset. Training three large deberta models with different datasets, and combining their features to inferring the right answer.

For images generated from text using Stable Diffusion, we use three models BLIP+CLIP, OFA and ViT. Then, we combine their features to predict the text for a given generated image.

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.

As an incarnation of matter inhaled by the deep breath, you should try your best to avoid sinking into the human body.

"WASD" move; "R" recover "Oxygen"; "E" interact with environment; "left Shift" sprint; "left mouse button" attack; "right mouse button" parry.

Scale your device and escape from this geometry storm.

This game reached the "Innovation RK1" and "Theme interpretation RK2" at Game Off 2023