S3OD: Towards Generalizable Salient Object
Detection with Synthetic Data

Large-Scale Synthetic Dataset for SOD • Ambiguity-Aware Architecture • State-of-the-Art Model

Orest Kupyn Hirokatsu Kataoka Christian Rupprecht

University of Oxford, VGG • AIST

📄 arXiv Paper 🚀 Live Demo 💻 Code 🤗 Dataset Viewer

🎯 TL;DR

We present two key contributions: (1) S3OD Dataset — 139K+ high-resolution synthetic images generated through a multi-modal diffusion pipeline that extracts labels from FLUX DiT features, concept attention maps, and DINO-v3 representations, and (2) Ambiguity-Aware Architecture — a streamlined model with multi-mask decoder that naturally handles inherent ambiguity in salient object detection. Our approach unifies DIS and HR-SOD tasks, achieving state-of-the-art performance with strong cross-dataset generalization.

S3OD Overview: Synthetic dataset samples and model predictions

Dataset & Model Highlights

Addressing the data bottleneck through synthetic data generation

139K+
Synthetic Images
2× larger than all existing SOD datasets combined
SOTA
Performance
State-of-the-art across DIS and HR-SOD benchmarks
1,676
Object Categories
Diverse scenes across multiple domains
Multi-Modal
Diffusion Pipeline
High-quality complex data with accurate annotations

The S3OD Dataset

Scroll through our diverse, high-quality synthetic samples

🎨 Multi-Modal Diffusion Pipeline

Our pipeline simultaneously generates images and masks by extracting multi-modal signals during diffusion:

  • FLUX DiT Features — Rich spatial understanding encoded during generation
  • Concept Attention Maps — Object-level focus from cross-attention layers
  • DINO-v3 Visual Features — Robust semantic representations from self-supervised learning

This ensures strong image-label alignment and high-quality annotations without teacher model bottlenecks.

🔄 Iterative Generation Framework

Our feedback-driven approach dynamically identifies model weaknesses and adapts the sampling distribution:

  • Performance Monitoring — Evaluate model on validation set to identify weak categories
  • Adaptive Sampling — Prioritize generation of challenging object categories
  • Continuous Improvement — Dataset quality improves iteratively as it grows

Unlike static methods, this enables targeted data generation where the model needs it most.

Sample
Animals
Sample
Camping
Sample
Vehicles
Sample
Cuisine
Sample
Sports
Sample
Furniture
Sample
Wildlife
Sample
Big Cats

← Scroll to explore more categories →

🔍 Explore in Dataset Viewer 💾 Download Dataset

Cross-Dataset Generalization

Synthetic pre-training resulting in strong real-world generalization

Method Training Data DAVIS-S Fm↑ HRSOD-TE Fm↑ DUTS-TE Fm↑ DUT-OMRON Fm↑
InSPyReNet DIS-5K .921 .891 .845 .713
BiRefNet DIS-5K .919 .887 .860 .744
MVANet DIS-5K .907 .902 .852 .711
S3OD (Ours) DIS-5K .951 .923 .902 .808
S3OD (Ours) S3OD Synthetic Only .970 .954 .937 .860

🚀 Try It Yourself!

Upload your own images and see S3OD in action with our interactive demo on HuggingFace Spaces.

🎯 Launch Interactive Demo

Get Started in Seconds

Simple Python API for state-of-the-art segmentation

# Install from GitHub
pip install git+https://github.com/KupynOrest/s3od.git
# Import and initialize
from s3od import BackgroundRemoval
from PIL import Image

# Initialize detector (automatically downloads model from HuggingFace)
detector = BackgroundRemoval()

# Load and process image
image = Image.open("your_image.jpg")
result = detector.remove_background(image)

# Save result with transparent background
result.rgba_image.save("output.png")

# Access predictions
best_mask = result.predicted_mask  # Best mask (H, W) numpy array
all_masks = result.all_masks       # All masks (N, H, W) numpy array
all_ious = result.all_ious         # IoU scores (N,) numpy array
# Load the S3OD dataset from HuggingFace
from datasets import load_dataset

dataset = load_dataset("okupyn/s3od_dataset")

# Access samples
for sample in dataset['train']:
    image = sample['image']
    mask = sample['mask']
    # Process your data...
💻 Browse Code on GitHub

Citation

If you find S3OD useful, please cite our work

@article{s3od2025,
  title={S3OD: Towards Generalizable Salient Object Detection with Synthetic Data},
  author={Kupyn, Orest and Kataoka, Hirokatsu and Rupprecht, Christian},
  journal={arXiv preprint arXiv:2510.21605},
  year={2025}
}