Skip to content

nitid Roadmap

Last reviewed: 2026-09-17.

nitid is a focused library for DETR-style vision models, released under Apache 2.0 with its pretrained weights. The project is focused on a small number of well-supported workflows: load a model by name, predict, track, train, validate, export, inspect results, and deploy against real video sources.

Product direction

The public API should stay compact:

from nitid import NITID

detector = NITID("model1s", task="detect")
segmenter = NITID("model1s", task="segment")
semantic = NITID("model1s", task="semantic")

Supported tasks are:

  • detect — D-FINE object detection.
  • segment — D-FINE instance segmentation.
  • semantic — dense semantic segmentation on D-FINE features.

The guiding principle is depth over breadth. A task belongs in nitid when it has a clear implementation, a normal user API, dataset support, validation metrics, export coverage where realistic, tests, and documentation.

In scope

  • Python API and CLI for prediction, tracking, training, validation, export, download, and model information.
  • Automatic download/preparation of supported official checkpoints.
  • Prediction on images, videos, folders, streams, screen capture, and numpy arrays.
  • Result helpers for annotated outputs, masks, semantic maps, JSON, YOLO TXT, CSV, pandas DataFrames, timings, and crops.
  • Fine-tuning and validation for detection, instance segmentation, and semantic segmentation.
  • Training features: AMP, EMA, resume, per-epoch metrics, callbacks, W&B, and MLflow.
  • Export to ONNX/OpenVINO for all supported tasks, and TorchScript/TensorRT for detection and instance segmentation.
  • RTSP camera input and annotated video output.
  • Tests and CI that protect supported user flows.

Not in scope by default

Area Current decision
Becoming a general YOLO model zoo Not planned unless a future optional adapter is justified by measured demand.
Distributed training/autobatch Valuable but not a differentiator yet.
Quantization and mobile runtimes Important for edge deployment, but each target needs a tested support plan.
Annotation tools and dataset doctors Prefer external specialist tools unless a tight nitid workflow emerges.
New tasks Require an implementation, metrics, export story, tests, docs, and a clear reason to belong here.

Current status

Available today:

  • Detection and instance segmentation with official checkpoint downloads.
  • Semantic segmentation training, validation, inference, and ONNX/OpenVINO export.
  • Tracking with optional tracker dependencies.
  • RTSP camera input, including authenticated cameras.
  • Developer tests for model construction, datasets, training, validation, export, and result containers.

Known gaps:

  • Semantic benchmark parity still needs larger public benchmark runs.
  • TensorRT coverage has not been completed for semantic segmentation.
  • Distributed training and automatic batch sizing are not implemented.
  • Quantized deployment is not implemented.
  • Public PyPI release workflow and hosted model registry remain release tasks.

Path to public release

nitid is not yet public. The remaining work to publish the repository and ship pip install nitid at v0.9.1 is tracked in the launch blockers milestone, which is release and publication work rather than features: third-party attribution, the published import/CLI name, wheel namespacing, branch and changelog sync, and the PyPI pipeline. The known gaps above are documented gaps for v0.9.1, not blockers.

Near-term priorities

  1. Stabilize the task surface: detection, instance segmentation, and semantic segmentation.
  2. Keep README/API/fine-tuning/export docs aligned with the actual public API.
  3. Add benchmark-grade validation runs for segmentation on datasets larger than smoke tests.
  4. Finish release hygiene: package artifacts, notices, versioning, and PyPI.
  5. Harden deployment workflows: reproducible export checks, hardware notes, and camera-runtime documentation.

Documentation