CVPR 2024论文阅读（资源整理【1】）
- 一、3d 重建与建模
-
- 论文1-Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction
- 论文2- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering
- 论文3-GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models
- 论文4-GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting
- 论文5-GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians
- 二、Vision Transformer和Transformer
-
- 论文1-TransNeXt: Robust Foveal Visual Perception for Vision Transformers
- 论文2-RepViT: Revisiting Mobile CNN From ViT Perspective
- 论文3-A General and Efficient Training for Transformer via Token Expansion
- 论文4-RMT: Retentive Networks Meet Vision Transformers
- 论文5- Learning Correlation Structures for Vision Transformers
- 论文6- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration
- 论文7 ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions
- 论文8 -Contextual Position Encoding: Learning to Count What's Important
- 论文9-Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers（AAAI2024）
- 论文10 - Fairness-Aware Structured Pruning in Transformers（AAAI2024）
- 论文11-SeTformer is What You Need for Vision and Language（AAAI2024）
- 论文12-iTransformer: Inverted Transformers Are Effective for Time Series Forecasting（ICLR 2024）
- 参考资料
一、3d 重建与建模
论文1-Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction
- Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction
- https://arxiv.org/abs/2309.13101
- https://github.com/ingra14m/Deformable-3D Gaussians
- 摘要: 隐式神经表示（Implicit Neural Representations）为动态场景重建和渲染的新方法铺平了道路。尽管如此，先进的动态神经渲染方法很大程度依赖这些隐式表示，经常难以捕捉场景中物体的复杂细节。此外，隐式方法在一般动态场景中实现实时渲染，限制了它们在各种任务中的使用。为了解决这一问题，我们提出了一种可变形的三维高斯分布的splatting 方法，该方法使用三维高斯分布来重建场景，并在具有变形场的规范空间中学习它们，以建模单目动态场景。我们还引入了一种没有额外开销的退火平滑训练机制，它可以减轻不准确的姿态对真实世界数据集中时