Practical overview
When it fits
Use it when you need a documented research implementation for transformer-based object detection pre-training and a path from ImageNet pre-training to COCO fine-tuning.
Who it is for
- Researchers studying unsupervised or self-supervised pre-training for object detection.
- Developers prototyping detection-transformer research workflows with PyTorch.
Documented uses
- Pre-train a detection transformer on ImageNet images without human annotations.
- Fine-tune the model on COCO 2017 for object detection.
- Inspect or build on the released pre-training and COCO fine-tuned model artifacts.
Requirements & limits
The repository is public and lists an Apache-2.0 license.
- The documented training examples use eight GPUs on a single node.
- The README estimates roughly five days for its 60-epoch pre-training example on eight V100 GPUs.
- The repository provides both a pre-training model and COCO fine-tuned weights.
- The project reports 43.1 box AP for its 300-epoch COCO fine-tuned model; the figure is project-reported.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare and run unsupervised pre-training
Run the repository’s documented 60-epoch UP-DETR pre-training workflow on one node with eight GPUs.
Before you start
- Python environment with PyTorch, torchvision, Cython, SciPy and the COCO API installed as described in the README.
- ILSVRC2012 training images downloaded and extracted into category directories.
- A single node with eight GPUs, as specified by the example command.
Steps
Install the listed dependencies using the README’s conda and pip instructions.
Prepare the ILSVRC2012 training images in the expected directory structure; image order is not required.
Run the documented distributed-launch command in the repository with the stated pre-training options, setting the ImageNet path and output directory.
Expected outcome
The command is configured to pre-train UP-DETR for 60 epochs and save the model to the specified output directory.
Guide limitations
- The README estimates about five days for 60 epochs on eight V100 GPUs; runtime can differ on other hardware.
- The guide describes the project’s setup instructions and does not verify a completed run.
Fine-tune UP-DETR on COCO 2017
Follow the documented eight-GPU, 300-epoch COCO fine-tuning setup.
Before you start
- The repository and dependencies installed as described in its README.
- COCO 2017 train and validation data, including annotations, organized in the expected directory structure.
- Eight GPUs for the README’s documented command.
Steps
Download and extract COCO 2017 train and validation data.
Arrange the files under the expected annotations, train2017 and val2017 directories.
Use the README’s distributed fine-tuning command with the COCO data path and pre-training checkpoint path set for the local setup.
Expected outcome
The documented command configures fine-tuning for 300 epochs using COCO 2017.
Guide limitations
- The excerpted instructions specify the data layout and command parameters but do not establish runtime or results on other hardware.
- This guide describes the repository instructions and is not a verification that fine-tuning completed successfully.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- dddzg/up-detr · Official homepage title · retained in its source language