Splatter Joint: 3D Gaussian Splatting for Articulated ObjectsIROS 2025 Oral

Junyan Li, Yifan Han, Pengfei Yi, Wenzhao Lian

To address the difficulty of existing methods in perceiving 3D articulated objects, we propose a novel approach that jointly trains 3D Gaussian Splatting (3DGS) with articulation parameters.

Specifically, the method takes a single-view image sequence as input, simultaneously reconstructing the 3DGS and predicting articulation parameters (e.g., rotation axes, opening angles). A specially designed 3DGS rendering pipeline then differentiably incorporates these parameters into 2D renderings, enabling supervision of the parameters directly from 2D images.

An early version of this work was selected as an outstanding undergraduate thesis at Shandong University and was ultimately accepted as an oral presentation at the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).

Paper    Code

Lane Detection with Deep Learning: Methods and DatasetsJun 2023, SCI Q4

To address the issue of excessive curvature and horizontal segments in suburban roads, I annotate data accordingly and refine the model.

Second Prize in 18th National College Student Smart Car CompetitionAug 2023

In the 18th National College Student Smart Car Competition Electric Relay Group, I was primarily responsible for software design for the charging section and communication and coordination between two cars.

I had to ensure not only the efficiency and stability of the charging process but also the precise data exchange and collaborative operation between the two vehicles. By writing and optimizing the charging algorithm, I ensured the efficient transfer of energy. At the same time, I designed a communication protocol that allowed the two cars to share critical information such as position, speed, and charging status in real-time, achieving seamless relay charging. Additionally, I was responsible for debugging and optimizing potential issues that might arise during the coordination process of the two cars, ensuring the smoothness and reliability of the entire charging process.

Second Prize in 17th National College Student Smart Car CompetitionAug 2022

I participated in the 17th National College Student Smart Car Competition in the iFlytek Smart Service Group, and independently completed the entire process from data collection, annotation to model training.

I also learned about neural network quantization and deployment techniques, and successfully deployed YoloV5 for object detection on the Jetson Nano platform. At the same time, I gained a preliminary understanding of the communication methods in the ROS system, which is crucial for the coordination and integration of different modules within the team.

Code

Second Prize in 17th National College Student Smart Car CompetitionAug 2022

I served as the team captain in the 17th National College Student Smart Car Competition, Balance Beacon Group, leading our team to secure the second prize in the category.

Within the team, I was primarily responsible for the task of beacon light detection using a camera.

I improved the beacon light detection algorithm that I developed last year to adapt to the car that could rock back and forth.

Video

Second Prize in 16th National College Student Smart Car Competition Aug 2021

I served as the team captain in the 16th National College Student Smart Car Competition, Energy-saving Beacon Group, leading our team to secure the second prize in the category.

Within the team, I was primarily responsible for the task of beacon light detection using a camera.

The main challenge of this task was to achieve real-time detection of small beacon lights at a distance on an embedded platform with limited computational capabilities, while mitigating interference caused by reflections and ambient lighting. I began by employing digital image processing techniques to filter, binarize, and detect edges in the images. Following this, I calculated the connected components within the images and extracted features such as size, coordinates, and rectangularity of these components. Subsequently, I fed these features into a multi-layer perceptron model to perform binary classification on each connected component. During the implementation, I manually derived the network and summarized patterns, developing a compact framework capable of stacking multiple fully connected layers for parameter tuning using C program language, which ultimately led to significant results.

Video1    Video2

Gold Medal in the CCF CCSP 2022Dec 2022

Gold Medal in the CCF CCSP (East China) 2022Dec 2022

Gold Medal in the CCF CCSP (East China) 2021Dec 2021