# Peterande/D-FINE

D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement  [ICLR 2025 Spotlight]

Repository: https://github.com/Peterande/D-FINE
Canonical: https://ross.abutalabs.com/products/d-fine
Language: Python
License: Apache-2.0
License Family: permissive
Topics: detr, object-detection, d-fine
Last push: 2026-08-19T15:08:14+00:00

## Health v2 (maintenance only)
Score: 67/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 98, release rhythm 35, longevity 53
- inputs: {"age_days": 751, "days_push": 14, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3305, forks 318 (observed 2026-08-28T04:07:55.838069+00:00)

## What it is
D-FINE is the official PyTorch implementation of an ICLR 2025 Spotlight paper that redefines the regression task in DETR-style detectors as fine-grained distribution refinement, achieving state-of-the-art real-time object detection on COCO. It provides pretrained models, training, and inference tooling for real-time object detection.

## Use cases
- train a state-of-the-art real-time object detector on custom datasets
- run fast object detection on images or video
- fine-tune a DETR-based detection model
- deploy an efficient object detection model with low latency
- reproduce ICLR 2025 D-FINE paper results on COCO
- export detection models for edge or production inference

## When to choose
- you need a fast, accurate real-time object detector with pretrained weights
- you want a modern DETR-family detector that is easy to train without hand-tuned anchors or NMS
- you need a research baseline for detection regression refinement

## When to avoid
- you need segmentation, tracking, or other vision tasks beyond bounding-box detection
- you lack GPU resources for training DETR-style models
- you need a detector integrated into a larger framework like MMDetection or Ultralytics

## Facets
- artifact type: library
- maturity: active
- function: computer-vision, machine-learning, deep-learning
- domain: computer-vision, deep-learning, artificial-intelligence
- platform: python, cross-platform
- tags: object-detection, detr, real-time-detection, iclr-2025, pytorch, model-zoo, gpu, linux

## Member repositories
- Peterande/D-FINE (main) score 67

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:55.838069+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:42:39.014614+00:00, confidence not recorded.
  - readme: https://github.com/Peterande/D-FINE (fetched 2026-08-28T04:07:55.838069+00:00, sha a2b8d55ad008)
- Data as of 2026-08-30T08:39:29.467469+00:00.