A Docker-powered microservice for PDF layout analysis, OCR, and content extraction that converts PDFs into structured Markdown or HTML with automatic translation support.

Street Gaussians is a research‑grade codebase that extends 3‑D Gaussian splatting to reconstruct and render dynamic street‑level scenes (e.g., Waymo data). It provides data‑conversion tools, training/evaluation scripts, and optional LiDAR‑depth/sky‑mask preprocessing, all wrapped in configurable YAML pipelines.

InfiniDepth is a CVPR‑2026 project that turns a single RGB image (or RGB + sparse depth) into arbitrary‑resolution depth maps and 3D Gaussian‑splatting models. It provides ready‑to‑run inference scripts, a Hugging Face Gradio demo, multi‑view video processing, and a full training pipeline for fine‑tuning on custom data.

EfficientLoFTR is a semi-dense local feature matching system that provides high-quality image correspondences with speeds comparable to sparse matching methods.

Neural Body is a research codebase that learns implicit neural fields with structured latent codes to model and render dynamic human bodies from multi‑view video. It supports training on ZJU‑MoCap and People‑Snapshot, rendering novel views, pose‑generalisation, mesh extraction, and provides Docker, distributed training, and pretrained models.

A framework for creating realistic, animatable 3D human avatars from videos using implicit neural representations like NeRF and SDF.

A minimalist Transformer-based framework for multi-class unsupervised anomaly detection that bridges the performance gap between unified multi-class models and single-class specialists.

An Android app that enables private, on-device semantic search for local photos using natural language descriptions or image queries via MobileCLIP2.

Scal3R is a scalable 3D reconstruction framework that uses test-time training to generate camera poses, depth maps, and point clouds from image sets.

A multi-view 3D reconstruction framework that combines SfM-guided monocular depth estimation with TSDF fusion to create accurate dense 3D models from images.

A collection of Google DeepMind's research projects on 4D representations, providing self-supervised video and vision models for spatial-temporal tasks like depth estimation and object tracking.

A tool for automatic 2D motion capture and biomechanical analysis that computes joint trajectories and angles from a single video or webcam.

AutoGaze is an autoregressive gazing model that prunes redundant video patches for vision transformers and multimodal LLMs, enabling efficient processing of high-resolution long-form videos. It is released by NVILabs with trained models, training data, a benchmark, and integration examples.