real-time object detection / instance segmentation
RF-DETR is a real-time DETR-style object detection, instance segmentation, and keypoint detection model architecture from Roboflow, using a DINOv2 backbone and intended to be fine-tuned on custom data.
Same problem, same approach. Swapping one for another is a config change, not a rewrite.
choose it for an actively maintained Apache-2.0 DETR-style detector with fine-grained distribution refinement when pure object detection is enough.
D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement [ICLR 2025 Spotlight]
choose it when you want an actively maintained real-time DETR detector with official code and detection-only COCO models.
[CVPR 2024] Official RT-DETR (RTDETR paddle pytorch), Real-Time DEtection TRansformer, DETRs Beat YOLOs on Real-time Object Detection. 🔥 🔥 🔥
Solves the same problem with a different architecture or at a different layer. Expect to rewrite the integration.
choose it when you want a broad OpenMMLab toolbox with many configurable detection/segmentation models rather than a single model package.
OpenMMLab Detection Toolbox and Benchmark
choose it when you are already in the PaddlePaddle ecosystem and need detection, segmentation, tracking, and pose models from one toolkit.
Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.
choose it when you want YOLO-family CNN models with an actively maintained training/export/deployment codebase.
Ultralytics YOLO26, YOLO11, YOLOv8 — object detection, instance segmentation, semantic segmentation, image classification, pose estimation, object tracking
choose it when you need open-vocabulary detection with arbitrary text-defined classes rather than fixed-class fine-tuning.
[CVPR 2024] Real-Time Open-Vocabulary Object Detection
choose it in PaddlePaddle projects when you want a YOLO/RT-DETR-family model zoo and associated training flows.
🚀🚀🚀 YOLO series of PaddlePaddle implementation, PP-YOLOE+, RT-DETR, YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv10, YOLO11, YOLOX, YOLOv5u, YOLOv7u, YOLOv6Lite, RTMDet and so on. 🚀🚀🚀
choose it when you are doing research on DETR-style detection and need a modular platform for building/benchmarking variants.
detrex is a research platform for DETR-based object detection, segmentation, pose estimation and other visual recognition tasks.
An earlier or more limited way to do the job. Still the right call when you need something small, proven, or CPU-only.
choose it only as the original DETR reference for research/education; it is old, unmaintained, and not real-time.
End-to-End Object Detection with Transformers
choose it if you need a simple keypoint-based detection/pose/3D baseline and accept an old, unmaintained codebase.
Object detection, 3D detection, and pose estimation using center point detection:
choose it if you need the archived universal mask-classification segmentation reference and do not need box detection.
Code release for "Masked-attention Mask Transformer for Universal Image Segmentation"
choose it for a box-free fully convolutional instance segmentation baseline, accepting old code and no maintenance.
SOLO and SOLOv2 for instance segmentation, ECCV 2020 & NeurIPS 2020.
choose it if you want the older unified transformer detection/segmentation baseline from CVPR 2023, despite being unmaintained.
[CVPR 2023] Official implementation of the paper "Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation"
choose it if you specifically need the 2020 original CenterMask implementation for real-time anchor-free instance segmentation.
[CVPR 2020] CenterMask : Real-Time Anchor-Free Instance Segmentation
choose it if you want the CenterMask method in a Detectron2-style implementation and can tolerate an old unmaintained codebase.
[CVPR 2020] CenterMask : Real-time Anchor-Free Instance Segmentation
choose it if you need the official DINO DETR baseline from ICLR 2023, accepting an older unmaintained repository.
[ICLR 2023] Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection"
choose it if you need an old unofficial TensorRT-oriented YOLOv7-with-transformers baseline, but expect no maintenance.
🔥🔥🔥🔥 (Earlier YOLOv7 not official one) YOLO with Transformers and Instance Segmentation, with TensorRT acceleration! 🔥🔥🔥
choose it for a fully convolutional one-stage anchor-free detector baseline if you do not need transformers and can use unmaintained code.
FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
These projects were analysed and named rf-detr among their alternatives. The relationship is not symmetric — how rf-detr rates them is a separate judgement, made when rf-detr is analysed in its own right.
calls rf-detr “Modern transformer-based detection/segmentation model; choose for real-time speed and COCO accuracy, but it is a specific architecture.”
Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
calls rf-detr “Choose if you want a maintained real-time DETR-style detector with detection+segmentation and Apache-2.0 code.”
Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
calls rf-detr “Choose it if you want a transformer-based real-time detector with strong fine-tuning performance and commercial-friendly license.”
YOLOv10: Real-Time End-to-End Object Detection [NeurIPS 2024]
calls rf-detr “Choose when you need an actively developed real-time DETR architecture for fine-tuning detection/segmentation; Apache-2.0 is permissive.”
Object detection, 3D detection, and pose estimation using center point detection:
calls rf-detr “Take it for Apache-2.0 real-time DETR backbones with segmentation when closed-set fine-tuning is enough.”
[CVPR 2024] Real-Time Open-Vocabulary Object Detection
calls rf-detr “Choose for a modern real-time DETR-based segmentation model with active maintenance and fine-tuning support.”
SOLO and SOLOv2 for instance segmentation, ECCV 2020 & NeurIPS 2020.
calls rf-detr “Specific real-time DETR-family model from Roboflow — choose for fine-tuning that architecture, not a full toolkit.”
Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.