[ICCV'25] St4RTrack: Simultaneous 4D Reconstruction and Tracking in the World 阅读报告
本文提出了一个 feed-forward 框架，通过引入一种创新的、依赖于时间的 pointmap 表示，并利用一个双分支 Transformer 架构，实现了在统一的世界坐标系中同时进行动态场景的密集追踪与三维重建。
Computer Vision
[ICCV'25 Oral] Back on Track: Bundle Adjustment for Dynamic Scene Reconstruction 阅读报告
本文巧妙地提出了一种“运动解耦”机制，通过一个学习的 3D Tracker 将动态物体的自身运动从观测运动中剥离，使得经典的 Bundle Adjustment 能够首次被统一地应用于含动态物体的场景中，极大地提升了动态场景重建中的相机位姿精度和三维重建质量。
[ICCV'25] SpatialTrackerV2: 3D Point Tracking Made Easy 阅读报告
本文提出了一个 feed-forward 3D point tracking architecture，它将 video depth、camera pose 和 object motion 进行统一建模和 end-to-end 优化，并通过在 17 个异构数据集上的可扩展训练，实现了 SOTA 的 3D 追踪精度和推理速度。
A Brief Exploration to Variational Autoencoder (VAE) with Code Implementation
Learn variational autoencoder (VAE) by reading and analyzing the paper: “Auto-Encoding Variational Bayes”. This post will introduce the basic work of VAE, including the derivation of formulas and simple code verification.
[NeurIPS'19 Oral] Generative Modeling by Estimating Gradients of the Data Distribution 阅读报告
This paper introduce a new generative model where samples are produced via Langevin dynamics using gradients of the data distribution estimated with score matching. And it is important to learn Score-Based generative network and Ito diffusion SDE.
[T-PAMI'23] Image Super-Resolution via Iterative Refinement 阅读报告
Image super-resolution with conditional diffusion model.
[CVPR'22] Deblurring via Stochastic Refinement 阅读报告
Image deblurring with “predict-and-refine” conditional diffusion model. An brand new strategy for ill-posed problem.
A Brief Exploration to Diffusion Probabilistic Models with Code Implementation
Learn diffusion probabilistic models (DPM) by reading and analyzing the papers: “Deep Unsupervised Learning using Nonequilibrium Thermodynamics” and “Denoising Diffusion Probabilistic Models”. This post will introduce the basic work of DPM, including the derivation of formulas and simple code verification.