Practical overview
When it fits
Use Vistr when you need a research implementation and are prepared to manage its local dependencies, model weights and video data.
Who it is for
- Computer-vision researchers reproducing the VisTR paper or its YouTubeVIS experiments.
- Developers prototyping video instance segmentation with a locally run PyTorch model.
Documented uses
- Run inference on video using a VisTR checkpoint and save results as JSON.
- Train the baseline model with YouTubeVIS data and pretrained DETR weights.
- Compare the documented ResNet-50 and ResNet-101 baseline configurations.
Requirements & limits
The public GitHub repository is open source under the Apache-2.0 license.
- The documented stack uses Python with PyTorch 1.6.0 and torchvision 0.7.0.
- Compiling the DCN module requires GCC 5.3 or later and CUDA 10.0 or later.
- Training requires at least 32 GB of GPU memory; the described run uses eight GPUs.
- Training preparation requires 2019 YouTubeVIS data and pretrained DETR weights.
- The README lists baseline checkpoints for ResNet-50 and ResNet-101 and directs users to save inference output as JSON.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the documented environment and data
Set up the dependencies and dataset layout described for VisTR before training or inference.
Before you start
- A local clone of the repository.
- Conda, PyTorch 1.6.0, torchvision 0.7.0, Cython, SciPy and pycocotools.
- GCC 5.3 or later and CUDA 10.0 or later for compiling the DCN module.
- For training, the 2019 YouTubeVIS images and annotations and pretrained DETR weights.
Steps
Clone the repository locally and install PyTorch 1.6.0 with torchvision 0.7.0.
Install Cython, SciPy and the COCO and YouTubeVIS pycocotools dependencies using the README instructions.
Compile the DCN module from models/dcn with the documented GCC and CUDA prerequisites.
For training preparation, arrange YouTubeVIS train and validation data and annotations under the README's expected data structure, and save pretrained DETR weights to the pretrained path.
Expected outcome
A prepared local environment and, when the training data and pretrained weights are supplied, the documented inputs for the training workflow.
Guide limitations
- The README does not provide a single environment specification covering all dependencies.
- Training requires at least 32 GB of GPU memory; the described run uses eight GPUs.
- The documented dataset is the 2019 YouTubeVIS release.
Run inference with a VisTR model checkpoint
Use the documented inference command to write segmentation results to a JSON file.
Before you start
- A prepared VisTR local environment.
- A VisTR model checkpoint, such as one of the listed ResNet-50 or ResNet-101 weights.
Steps
Obtain a model checkpoint from the model links listed in the repository README.
Run the command below, replacing the placeholders with the local checkpoint and desired output paths.
python inference.py --masks --model_path /path/to/model_weights --save_path /path/to/results.json
Expected outcome
The inference workflow is configured to save results to the specified JSON path.
Guide limitations
- The README command requires local model weights and does not document a hosted inference service.
- The reported inference-time figure excludes data loading and post-processing.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- YuqingWang1029/VisTR · Official homepage title · retained in its source language