Youtian Lin 林尤添

I am a first-year Ph.D. student at Nanjing University, supervised by Prof. Yao Yao and Prof. Tieniu Tan. My research focuses on 3D/4D reconstruction and generation. Previously, I pursued a Ph.D. at the Harbin Institute of Technology. I earned my M.S. from the Harbin Engineering University in 2021, where I was advised by Prof. Jian Guan.

Email  |  Scholar  |  Twitter (X)  |  Github

News

2024-09-27: Two papers accepted by NeurIPS 2024 ! 🎉🎉🎉 Direct3D, FastDrag

2024-07-01: Three papers accepted by ECCV 2024 ! 🎉🎉🎉 Relightable 3D Gaussian, STAG4D, UniDream

2024-05-24: New paper released! FastDrag

2024-05-23: New paper released! Direct3D

Research

A method for distilling geometric information from a pre-trained optical flow model into 3DGS. FDS samples unobserved views adjacent to the input views and calculates Prior-Flow to guide the analytically calculated Radiance-Flow.

Direct3D introduces a scalable approach for generating high-quality 3D assets from images. It uses D3D-VAE for efficient 3D shape encoding and D3D-DiT for modeling 3D latents. This method setting a new standard for 3D content creation.

FastDrag is a new, faster drag-based image editing method that uses a latent warpage function for one-step pixel adjustment and a bilateral nearest neighbor interpolation to fill null regions. It also ensures consistency with the original image.

High-fidelity 4D generation from diverse inputs (text, image, and video) with pre-trained diffusion models and dynamic 3D Gaussian splatting.

We propose an innovative point-based method for rapid dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos, leveraging advancements in point-based 3D Gaussian Splatting (3DGS).

Utilizuing 3D Gaussian points to represent a scene, allowing for material and lighting decomposition, enabling real-time relighting, ray-tracing, and editing of the 3D point cloud with improved BRDF estimation and novel view rendering results.

Use a dual-phase training process for albedo-normal aligned multi-view diffusion and reconstruction models, a progressive generation procedure for geometry and albedo-textures using Score Distillation Sample (SDS), and an innovative SDS application for finalizing Physically Based Rendering (PBR) generation with fixed albedo.

We extend the Instant-NGP framework to support dynamic scenes, and show that it can be used to train a dynamic NeRF model that is both more compact and more efficient than prior work.

Incorporating adaptive scale sampling, dynamic elliptical distribution aided sampling, and spatial distance weighting to enhance the selection of high-quality positive samples.

Utilizing a one-stage keypoint based network architecture and introducing a novel geometric transformation method to achieve orientation angle regression, along with incorporating Student's-t distribution to enhance performance

We addressing the challenges of computational complexity in two-stage detectors by employing a per-pixel prediction approach with a geometric transformation, a branch interactive module, and an enhanced intersection over union (IoU) loss.

Project

pointrix-project
Pointrix: a differentiable point-based rendering library.

threestudio-3dgs
The Gaussian Splatting extension for threestudio.

taichi-nerfs
A PyTorch + Taichi implementation of instant-ngp NeRF training pipeline. For more details about modeling, please checkout this blog.

taichi-ngp-renderer
A Instant-NGP renderer implemented using Taichi, written entirely in Python. No CUDA!

nerfacc
A PyTorch Nerf acceleration toolbox for both training and inference. (As a contributor, I have implemented a fast ngp rendering in CUDA)

Source code borrow from jonbarron.