|
I am a first-year Ph.D student in the Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University , advised by Prof. Rongrong Ji. My research interest lies in the Machine Learning and 3D Computer Vision. Email / CV / Google Scholar / Github |
|
|
|
|
* indicates equal contribution |
|
Yansong Qu*, Shaohui Dai* , Xinyang Li , Jianghang Lin , Liujuan Cao, Shengchuan Zhang, Rongrong Ji ACMMM, 2024 [Project Page] [arXiv] [Code] In this paper, we propose GOI, a novel method for 3D open-vocabulary scene understanding. Our approach includes an efficient compression method that utilizes scene priors to condense noisy high-dimensional semantic features into compact low-dimensional vectors, which are subsequently embedded in 3DGS. And we leverage an off-the-shelf 2D Referring Expression Segmentation (RES) model to fine-tune the semantic-space hyperplane, enabling a more precise distinction between target regions and others. |
|
Yansong Qu*, Yuze Wang*, Yue Qi ICME, 2023 [Paper] [Code] Neural Radiance Fields (NeRF) require a large number of high-quality images to achieve novel view synthesis in room-scale scenes, but capturing these images is very labor-intensive. To address this, we propose Semantic-Guided Point-Based NeRF (SG-NeRF), which can reconstruct the radiance field using only a few images. We leverage sparse 3D point clouds with neural features as geometry constraints for NeRF optimization and use semantic predictions from both 2D images and 3D point clouds to guide the search for neighboring neural points during ray marching. This semantic guidance allows the sampled points to accurately find structurally related points even in large areas with unevenly distributed sparse point clouds, enabling high-quality rendering with fewer input images. |
|
Yuze Wang, Junyi Wang, Yansong Qu, Yue Qi ICMR, 2023 [Paper] In this work, we introduce Rotation-Invariant Point-based NeRF (RIP-NeRF), combining the strengths of implicit NeRF-based and explicit point-based representations for fine-grained editing and cross-scene compositing of radiance fields. We replace the traditional Cartesian coordinates with a novel rotation-invariant point-based radiance field representation, using a Neural Inverse Distance Weighting Interpolation (NIDWI) module to enhance rendering quality. For cross-scene compositing, we disentangle the rendering module from the neural point-based representation, allowing controllable compositing across scenes without the need for retraining. |
|
Conference Review:Journal Review: |
|
|
|
|