OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Selfreformer

Selfreformer is the official Python research implementation of SelfReformer, a self-refined deep-learning network for salient object detection that aims to identify salient objects while preserving image detail. Its GitHub repository documents training and testing workflows, dataset organization, checkpoint use and evaluation.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Selfreformer — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Selfreformer when a research workflow needs the SelfReformer implementation for training or evaluating salient-object predictions, and the documented Python, PyTorch and GPU requirements fit the available environment.

Who it is for

  • Computer-vision researchers studying salient object detection.
  • Practitioners who want to reproduce the SelfReformer research implementation.
  • Researchers comparing saliency models or adapting this network for experiments.

Documented uses

  • Train the model on a salient-object-detection dataset.
  • Test a pretrained checkpoint on a benchmark dataset.
  • Save saliency prediction maps for subsequent evaluation.
  • Organize datasets and benchmark folders in the layout used by the repository.
Before you choose

Requirements & limits

The implementation is a public GitHub repository marked BSD-3-Clause.

  • The documented environment specifies Ubuntu 18.04, Python 3.8.3, Torch 1.8.2+cu111, Torchvision 0.9.2+cu111 and Kornia.
  • Git LFS is needed for large files; the README also provides a Google Drive download if the LFS file cannot be obtained.
  • The default training batch size of 16 is stated to require 24 GB of GPU RAM; reducing batch size is suggested for smaller hardware.
  • The README says training currently supports a single GPU.
  • Evaluation metrics use separate tools referenced by the repository, including code from BASNet and F3Net.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare datasets for SelfReformer training or evaluation

Organize image and mask files in the directory structure described for the model.

Before you start

  • A dataset with image files and corresponding mask files.
  • For training, a dataset folder placed under dataset/.
  • For evaluation, benchmark datasets placed under dataset/benchmark/.

Steps

  1. Create a folder for each dataset.

  2. Inside each dataset folder, create Images and Masks folders.

  3. Place image files such as .jpg files in Images and mask files such as .png files in Masks.

  4. Put training datasets under dataset/ and evaluation datasets under dataset/benchmark/.

Expected outcome

The dataset follows the folder layout expected by the repository's documented training and evaluation workflow.

Guide limitations

  • This guide covers dataset organization only; it does not verify dataset contents or run training or evaluation.
Documentation 1 ↗
Source-based practical guide

Set up the Swin backbone files

Place the released Swin backbone in the repository location specified by the README.

Before you start

  • A copy of the SelfReformer repository.
  • The Swin backbone file available through the README's linked download.

Steps

  1. Obtain the Swin backbone from the link in the repository README.

  2. Place the backbone under the model/pretrain folder.

Expected outcome

The Swin backbone file is located in the directory specified by the project instructions.

Guide limitations

  • The README does not provide further steps here for running or evaluating the Swin variant.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
16 October 2026
Original website title
BarCodeReader/SelfReformer · Official homepage title · retained in its source language
How we review sourcesSuggest a correction