Practical overview
When it fits
Use it for self-hosted research experiments on few-shot object detection, especially when you need to study adaptation to novel classes or reproduce the paper’s workflow.
Who it is for
- Computer-vision researchers reproducing the Meta-DETR paper’s results
- Practitioners studying object detection with limited labeled examples for new categories
- Teams benchmarking few-shot detection methods on COCO or Pascal VOC
Documented uses
- Reproduce the paper’s reported experiments using the provided scripts
- Train a base detector and fine-tune it with few-shot examples
- Run inference using a checkpoint produced by few-shot fine-tuning
- Set up COCO or Pascal VOC few-shot data for experiments
Requirements & limits
The repository is public and MIT-licensed. It is run locally rather than offered as a hosted detection service.
- The README says NVIDIA GPUs are required; its recommended configuration uses eight V100 GPUs.
- The implementation depends on CUDA-compiled Deformable Attention operators.
- COCO 2017 must be downloaded separately; sampled COCO and Pascal VOC few-shot setups are provided through the repository.
- The README links pretrained base-training weights for fine-tuning.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the Meta Detr code environment
Clone the repository, install its listed Python dependencies and compile the required Deformable Attention CUDA operators.
Before you start
- Linux with an NVIDIA GPU
- Python 3.7 and a CUDA-matched PyTorch/TorchVision installation; the README recommends CUDA 10.2, PyTorch 1.7.1 and TorchVision 0.8.2
- GCC and the listed dependencies: Cython, pycocotools, tqdm and SciPy
Steps
Clone the Meta-DETR repository and create and activate a Python 3.7 Conda environment.
Install PyTorch and TorchVision, matching the CUDA toolkit to the local environment.
Install Cython, SciPy, tqdm and the COCO API Python package.
From the repository, run the provided make script in models/ops to compile Deformable Attention CUDA operators.
Expected outcome
The documented code environment and CUDA operators are prepared for the project’s later data and experiment stages.
Guide limitations
- The repository requires NVIDIA GPUs to run its code.
- The README recommends a specific setup and lists other Linux and package-version ranges that may work; it does not guarantee compatibility.
- This setup guide does not cover dataset preparation or running training and inference.
Prepare MS-COCO data for few-shot experiments
Arrange COCO 2017 and the repository’s sampled few-shot sets in the directory layout expected by the project.
Before you start
- The Meta-DETR repository
- The COCO 2017 dataset, downloaded separately
Steps
Download COCO 2017.
Place its training images, validation images and annotations under data/coco, following the README’s directory layout.
Use the repository-provided coco_fewshot folder under data for the sampled few-shot setups.
Expected outcome
The documented COCO data folders are arranged for the project’s few-shot workflow.
Guide limitations
- The README describes five random-seed setups with 1, 3, 5, 10 or 30 instances per novel class; base-class instance counts vary.
- This guide covers data organization, not model training or evaluation.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- ZhangGongjie/Meta-DETR · Official homepage title · retained in its source language