Open Source Python Object Detection Models for ChromeOS

Frigate - NVR With Realtime Object Detection for IP Cameras A complete and local NVR designed for Home Assistant with AI object detection. Uses OpenCV and Tensorflow to perform realtime object detection locally for IP cameras. Use of a Google Coral Accelerator is optional, but highly recommended. The Coral will outperform even the best CPUs and can process 100+ FPS with very little overhead.

Powerful command line and configuration file parsing for C++, Python, Ruby and Java (others to come). This tool provides many features, such as separate treatment for options, variables, and flags, unrecognized object detection, prefixes and much more.

CutLER is an approach for unsupervised object detection and instance segmentation that trains detectors without human-annotated labels, and the repo also includes VideoCutLER for unsupervised video instance segmentation. The method follows a “Cut-and-LEaRn” recipe: bootstrap object proposals, refine them iteratively, and train detection/segmentation heads to discover objects across diverse datasets. The codebase provides training and inference scripts, model configs, and references to benchmarking results that report large gains over prior unsupervised baselines. It’s intended for researchers exploring self-supervised and unsupervised recognition, offering a practical path to scale beyond costly labeled corpora. The README links papers and gives a high-level overview of components and expected outputs, with pointers to demos and assets. The repository is actively starred and structured as a typical research release with license, contribution guidelines, and security policy.

DetectAndTrack is the reference implementation for the CVPR 2018 paper “Detect-and-Track: Efficient Pose Estimation in Videos,” focusing on human keypoint detection and tracking across video frames. The system combines per-frame pose detection with a tracking mechanism to maintain identities over time, enabling efficient multi-person pose estimation in video. Code and instructions are organized to replicate paper results and to serve as a starting point for researchers working on pose in video. Although the repo has been archived and is now read-only, its issue tracker and artifacts remain useful for understanding implementation details and experimental settings. The project sits alongside other Facebook Research vision efforts, offering historical context for the evolution of video pose and tracking techniques. Researchers can still study the algorithms, adapt the pipeline, or port ideas into modern frameworks.

Detic (“Detecting Twenty-thousand Classes using Image-level Supervision”) is a large-vocabulary object detector that scales beyond fully annotated datasets by leveraging image-level labels. It decouples localization from classification, training a strong box localizer on standard detection data while learning classifiers from weak supervision and large image-tag corpora. A shared region proposal backbone feeds a flexible classification head that can expand to tens of thousands of categories without exhaustive box annotations. The system supports zero- or few-shot extension to novel categories via semantic embeddings and class name supervision, making “open-world” detection practical. Built on Detectron2, the repo includes configs, pretrained weights, and conversion tools to mix fully and weakly supervised sources. Detic is especially useful for applications where label space is vast and long-tailed, but dense bounding-box annotation is infeasible.

VoteNet is a 3D object detection framework for point clouds that combines deep point set networks with a Hough voting mechanism to localize and classify objects in 3D space. It tackles the challenge that object centroids in 3D scenes often don’t lie on any input surface point by having each point “vote” for potential object centers; these votes are then clustered to propose object hypotheses. Once cluster centers are formed, the network regresses bounding boxes around them and classifies them. VoteNet works end-to-end: it learns the voting, aggregation, and bounding-box regression components jointly, enabling strong detection accuracy without relying on 2D proxies or voxelization. The codebase includes data preparation for indoor datasets (SUN RGB-D, ScanNet), training and evaluation scripts, and demo utilities to visualize predicted boxes over point clouds.