ABot-Recon

LingBot-Map is a feed-forward 3D foundation model from the Robbyant Team for streaming 3D reconstruction from image sequences and video. Its Geometric Context Transformer unifies coordinate grounding, dense geometric cues, and long-range drift correction through anchor context, a pose-reference window, and trajectory memory. The system uses paged KV-cache attention for streaming inference and supports interactive browser-based reconstruction with point-cloud visualization, keyframe intervals, windowed processing for long sequences, and sky masking. It also provides an offline rendering pipeline for image folders or video, along with evaluation pipelines for datasets including KITTI and Oxford Spires. The repository reports approximately 20 FPS inference at 518×378 resolution on sequences exceeding 10,000 frames, and documents PyTorch installation with FlashInfer as the recommended attention backend and native SDPA as a fallback.

VidMap is an offline structure-from-motion system for video, developed by the Computer Vision and Geometry group (cvg). It combines temporal feature tracks, loop closures, metric depth, and global optimization to estimate camera poses, camera intrinsics, and a sparse 3D map from video or extracted frames. The Python package writes the reconstruction as a COLMAP model and supports optional browser-based HTML visualization and Rerun playback or flythrough recordings. The repository provides setup and execution workflows built around COLMAP, PyCOLMAP, PyTorch, xFormers, and GPU acceleration through CUDA and cuDSS.

4DAnyone is an open-source generative AI pipeline from ant-research that converts a casual monocular video of one person into synchronized videos from multiple virtual camera views for downstream 3D or 4D Gaussian-splatting reconstruction. It supports configurable camera layouts, including full 360-degree or frontal-arc coverage, multiple pitch layers, adjustable yaw ranges, and flexible view counts. The repository provides inference scripts, automatically downloadable models and examples, motion-recovery outputs, camera metadata, skeleton videos, and generated target-view videos. Input footage should be at least 720p, preferably 1080p, use a 9:16 portrait aspect ratio, contain a single person in a full-body or upper-body shot, include at least 121 frames, and have only mild camera motion.

Rerun is an open-source SDK and data layer for recording, querying, visualizing, transforming, and streaming multimodal physical-robotics data. It ingests multi-rate data such as images, point clouds, transforms, time series, joint states, tensors, and video from robot logs, human-data rigs, simulation, and web video, using formats including MCAP, RRD, and LeRobot. Its built-in viewer synchronizes and renders the data in real time for episode scrubbing, sensor comparison, and live computer-vision pipelines; the same data can be queried with dataframes or SQL and streamed directly into training. Rerun is built in Rust on column-chunk storage designed for multi-rate physical data, with Python, Rust, and C++ SDKs and a separate Viewer binary for network streaming and RRD files. The repository describes the project as being in active development, with an evolving API and possible breaking changes.

FoundationPose is an NVIDIA Research implementation of a unified foundation model for 6D object pose estimation and tracking of novel objects. It supports model-based operation from a CAD model and model-free operation from a small set of reference images, applying to a new object at test time without fine-tuning. A neural implicit representation provides novel-view synthesis while keeping the downstream pose-estimation modules shared across both setups; the system uses large-scale synthetic training, a transformer-based architecture, contrastive learning, and language-model assistance. The repository includes robotic and augmented-reality demos and provides Docker-based setup instructions and downloadable model weights.