ParkerLabel

Overview

GitHub

ParkerLabel is a Python-based AI-assisted interactive image annotation tool designed primarily for object detection and instance segmentation. Its core feature is that it runs entirely on the CPU, requiring no GPU support. On CPUs such as the Apple M5 Pro and Intel Core i9-14900KF, each interactive mask-decoder inference takes about 20 ms. ### Key Capabilities - **AI-Assisted Annotation**: In AI mode, users guide the MobileSAM model with positive and negative clicks to generate masks for the current object. - **Multi-Mode Switching**: Offers three working modes—AI, Manual, and Inspect—with default shortcuts Q, W, and E respectively. - **Category Configuration Management**: Includes a built-in read-only configuration of 80 standard COCO categories. Users can create an empty configuration or duplicate an existing one to generate a new version. Saved configurations are immutable; modifications require duplication to produce a new version. - **Data Integrity Verification**: Each annotation is bound to a specific category configuration via UUID and SHA-256 digest. If the active configuration doesn't match, the app prompts you to switch; if the configuration is missing or verification fails, annotations open in read-only mode. - **Multilingual Interface**: Supports Simplified Chinese, Traditional Chinese, English, Japanese, Korean, German, French, Italian, and Spanish. ### File & Storage Formats Each source image corresponds to three adjacent files: - `.npy`: Stores reusable image embeddings. - `.json`: Stores the category, bounding box, area, and uncompressed COCO RLE mask at source resolution for each instance. This JSON uses COCO annotation fields but is not a complete COCO dataset file. - `.mask.png`: Stores a flattened color preview at source resolution. ### Deployment Portable release packages are available for macOS arm64 and Windows x64. You can also run from source by creating a Python 3.11 environment via Conda. The app automatically verifies and downloads missing ONNX models to the `pretrain/` directory. The project is open-sourced under the GPL-3.0 license.

Q&A

Does ParkerLabel require a GPU to run?AI-generated

No. ParkerLabel runs entirely on the CPU. On Apple M5 Pro and Intel Core i9-14900KF CPUs, each interactive mask-decoder inference takes only about 20 ms.

Which operating systems are supported?AI-generated

Portable release packages are available for macOS arm64 and Windows x64. You can also run from source on any supported platform using a Conda environment.

What format are annotations saved in?AI-generated

Three files are generated per source image: .npy stores image embeddings, .json stores the category, bounding box, area, and uncompressed COCO RLE mask for each instance, and .mask.png stores a color preview. The JSON uses COCO annotation fields but is not a complete COCO dataset file.

What should I do if macOS blocks the app on first launch?AI-generated

The macOS package is ad-hoc signed but not notarized by Apple. After your first attempt to open it, go to System Settings → Privacy & Security, scroll to the Security section, and click "Open Anyway".

How do I change the interface language?AI-generated

Change the interface language in-app via Settings → Interface Language. ParkerLabel supports Simplified Chinese, Traditional Chinese, English, Japanese, Korean, German, French, Italian, and Spanish.

Can category configurations be modified?AI-generated

Saved configurations are immutable. To make changes, you must duplicate an existing configuration (including the built-in COCO config) into a new version, then edit the new version.