# computer vision

This model for non-visible object segmentation and content completion in videos was proposed by Carnegie Mellon University. It processes sequences of visible objects in a video using a conditional generation approach, leveraging foundational knowledge from video generation models to produce masks and RGB content that include both visible and occluded parts. The main advantages of this technology include its ability to effectively handle highly occluded situations and deformable objects. Additionally, the model outperforms existing state-of-the-art methods across multiple datasets, showing up to a 13% performance improvement in the segmentation of non-visible areas obstructed by objects.

VGGSfM is a depth learning-driven 3D reconstruction technology aimed at reconstructing the camera pose and 3D structure of a scene from a set of unconstrained 2D images. This technology uses a fully differentiable deep learning framework for end-to-end training. It extracts reliable pixel-level trajectories using depth 2D point tracking technology, while restoring all cameras based on image and trajectory features, and optimizing camera and triangulated 3D points through a differentiable bundle adjustment layer. VGGSfM achieves state-of-the-art performance in three popular datasets: CO3D, IMC Phototourism, and ETH3D.

VastGaussian is an open-source project for 3D scene reconstruction that leverages 3D Gaussians to model the geometric and appearance information of large-scale scenes. This project is an implementation from scratch and may contain some errors, but it offers a new perspective for the field of 3D scene reconstruction. Key advantages include its ability to process large datasets and the improvements made to the original 3DGS project, enhancing its clarity and usability.

JavaVision is a comprehensive visual intelligent recognition project based on Java development. It not only achieves core functions such as PaddleOCR-V4 text recognition, YoloV8 object recognition, face recognition, and image search by image, but also can easily expand to other fields such as voice recognition, animal recognition, and security checks. The features of the project include the use of the SpringBoot framework, multifunctionality, high performance, reliability, easy integration, and flexibility. JavaVision aims to provide Java developers with a comprehensive visual intelligent recognition solution, allowing them to build advanced, reliable, and easily integrated AI applications with the familiar and beloved programming language.

FineControlNet is an official PyTorch implementation for generating images controlled by spatial-aligned text inputs, such as 2D human poses, and instance-specific text descriptions. It can handle a wide range of spatial inputs from simple line drawings to complex human poses. FineControlNet ensures natural interaction and visual coherence between instances and the environment, while retaining the high quality and generalization capabilities of Stable Diffusion, but with enhanced control.

ODIN (Omni-Dimensional INstance segmentation) is a model that uses a transformer architecture for segmentation and labeling on both 2D RGB images and 3D point clouds. It distinguishes 2D and 3D feature operations by iteratively fusing information between 2D views and 3D views. ODIN achieves state-of-the-art performance on ScanNet200, Matterport3D, and AI2THOR 3D instance segmentation benchmarks, and achieves competitive performance on ScanNet, S3DIS, and COCO. When using sampled point clouds from 3D meshes instead of perceived 3D point clouds, it surpasses all previous works. As the 3D perception engine in a guided concretization agent architecture, it sets a new state-of-the-art on the TEACh dialogue action benchmark. Our code and checkpoints can be found on the project website.

Roboflow is a comprehensive platform for building and deploying computer vision models. Used by over 250,000 engineers, it enables the creation of datasets, model training, and deployment to production environments. With Roboflow, you can train a working, state-of-the-art computer vision model in under 24 hours using just a few dozen example images. It provides a suite of functionalities including dataset management, annotation tools, model training, model deployment, and supports integration with various environments and tools.