Practical overview
When it fits
Use UPSNet when a local computer-vision research workflow needs panoptic segmentation experiments on COCO or Cityscapes and can accommodate the repository's dataset preparation and compute requirements.
Who it is for
- Computer-vision researchers evaluating panoptic segmentation on COCO or Cityscapes.
- Teams preparing those datasets for model training or testing.
- Developers building a research evaluation workflow around the supplied scripts and experiment configurations.
Documented uses
- Run image-based panoptic segmentation with the supplied COCO or Cityscapes weights.
- Train or fine-tune the model on prepared COCO or Cityscapes data.
- Evaluate model variants through the repository's dataset-specific configurations and test script.
Requirements & limits
The repository states that it is licensed under the Uber Non-Commercial License.
- The README reports Python 3.6 and PyTorch 0.4.1 as the tested software versions.
- The initialization process builds C++/CUDA modules.
- Training is recommended on 4–16 GPUs with at least 11 GB memory; the README reports 16-GPU training with Horovod.
- COCO and Cityscapes must be downloaded and prepared separately, with specific label-preparation requirements for Cityscapes.
- The license is non-commercial.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run the supplied COCO or Cityscapes test workflow
Prepare the repository and dataset, then use a supplied weight file and configuration with the test script.
Before you start
- A Linux environment with Python 3.6 and PyTorch 0.4.1, the README's stated tested setup.
- The repository and the relevant downloaded COCO or Cityscapes dataset.
- For Cityscapes, generated TrainIds label images; for COCO, images and annotations folders.
- The README's initialization scripts build C++/CUDA modules and prepare the dataset.
Steps
Clone the repository and run init.sh to build essential C++/CUDA modules and download a pretrained model.
For Cityscapes, link the dataset at data/cityscapes and run init_cityscapes.sh; for COCO, link it at data/coco and run init_coco.sh.
Run download_weights.sh to obtain the trained COCO and Cityscapes weights.
Run upsnet/upsnet_end2end_test.py with a matching experiment configuration and weight_path, using one of the dataset-specific examples in the README.
Expected outcome
The documented test workflow runs UPSNet evaluation using a supplied dataset-specific configuration and trained weights.
Guide limitations
- Dataset files must be downloaded separately and meet the stated preparation requirements.
- The README's tested software versions are Python 3.6 and PyTorch 0.4.1; it suggests other versions may work but does not guarantee compatibility.
- This is a repository workflow, not a hosted inference service.
Prepare a configured training run
Prepare COCO or Cityscapes data and select a supplied experiment configuration for model training.
Before you start
- Python and PyTorch; the README reports testing with Python 3.6 and PyTorch 0.4.1.
- Downloaded COCO or Cityscapes data prepared as specified in the README.
- The README recommends 4–16 GPUs with at least 11 GB memory and reports training with 16 GPUs using Horovod.
Steps
Clone the repository and run init.sh to build essential C++/CUDA modules.
Link the prepared dataset under data/coco or data/cityscapes and run its corresponding initialization script.
Choose a matching YAML configuration from upsnet/experiments.
Use the documented training entry point, upsnet/upsnet_end2end_train.py, with the selected configuration.
Expected outcome
The repository is prepared to start a training run using a supplied dataset configuration.
Guide limitations
- The README recommends substantial multi-GPU hardware for training.
- This guide covers the documented setup and invocation, not completion or measured results of a run.
- The project is licensed under the Uber Non-Commercial License.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- uber-research/UPSNet · Official homepage title · retained in its source language