hustvl

Professional software vendor delivering innovative solutions on the Softono platform. Specialized in both open-source and proprietary software development.

Visit Website

Total Products

Software by hustvl

Open Source

4DGaussians

4DGaussians is an open-source implementation of 4D Gaussian Splatting, a technique presented at CVPR 2024 for real-time dynamic scene rendering. Developed by researchers from HUST and Huawei, this software extends the original 3D Gaussian Splatting method to handle time-varying scenes with extreme efficiency. The method achieves real-time rendering speeds while maintaining high visual fidelity, with optimized training times of approximately 8 minutes for D-NeRF datasets and 30 minutes for HyperNeRF datasets. Key capabilities include fast convergence and support for both synthetic and complex real-world dynamic scenes. The project provides complete code for training and rendering, along with installation guides for dependency management including PyTorch, CUDA, and specific rasterization submodules. It supports multiple benchmark datasets such as D-NeRF, HyperNeRF, DyNeRF, and the Plenoptic Dataset. Users can visualize results using the SIBR viewer or utilize provided Colab demos for quick testing without loca

ML Frameworks 3D Modeling & Animation

3.7K Github Stars

Open Source

YOLOS

YOLOS stands for You Only Look at One Sequence, a research software project and machine learning model introduced at NeurIPS 2021 that rethinks object detection using Vision Transformers. The primary goal of YOLOS is not to create the highest performing detector available, but to demonstrate the versatility and transferability of vanilla Vision Transformers pre-trained on mid-sized ImageNet-1k to the challenging COCO object detection benchmark. By adding a single special detection token to the input sequence, the model performs pure sequence-to-sequence object detection using fixed-size non-overlapping image patches with minimal 2D inductive biases. This approach allows the system to detect objects in any dimensional space without explicitly understanding spatial dimensions beyond the sequence of patches. The project highlights that complex visual tasks traditionally requiring heavy convolutional biases can be achieved effectively through simple Transformer mechanisms. YOLOS has been integrated into the Huggi

ML Frameworks

903 Github Stars