IRIS Dataset: Industrial Real-Sim Imagery Set

Overview

The IRIS Dataset is a comprehensive real-world dataset designed to study sim-to-real transfer for object detection in industrial robotic environments. This repository provides:

1. The complete real IRIS dataset: 508 annotated images of 32 mechanical components captured across four distinct, challenging industrial scenes.
2. Assets for synthetic data generation: All necessary 3D models, backgrounds, and materials to run the companion synthetic data generation pipeline.
3. Example synthetic datasets: Two fully-annotated synthetic training sets (4000 images each) generated using our pipeline, showcasing different data generation strategies.
4. Pre-trained model checkpoints: YOLO11m models trained on the provided synthetic datasets, serving as baselines for sim-to-real transfer experiments.

This release accompanies our paper and the open-source synthetic data generation code SynthRender. The goal is to provide a complete, reproducible benchmark for evaluating and advancing sim-to-real methods in industrial robotics.

Dataset Statistics

TOTAL DATA: 508 images, 32 classes

Distribution by instance count:

• 96 single object images
• 210 single instance images
• 202 double instance images

Scene Breakdown:

Scene Type	Count	Image Range
Controlled lighting (room)	101	000–100
Window sunlight	67	101–167
Background diversity	100	168–267
Industrial robot scene	240	268–507

Folder Structure

IRIS
├── Assets
│   ├── CADs
│   │   ├── 3DGS
│   │   ├── Manual
│   │   ├── MeshyAI
│   │   └── TRELLIS
│   ├── General
│   │   ├── backgrounds
│   │   ├── distractors
│   │   └── plane_materials
│   └── 3D_GenAI_Masked_Imgs
├── Checkpoints
├── Real_Test_Set
│   ├── annotations
│   │   ├── coco
│   │   │   ├── by_scene
│   │   │   └── full
│   │   └── yolo
│   │       ├── by_scene
│   │       │   ├── 01_control_lighting
│   │       │   ├── 02_sunlight_window
│   │       │   ├── 03_floor_backgrounds
│   │       │   └── 04_robot_scene
│   │       └── full
│   └── images
│       ├── by_scene
│       │   ├── 01_control_lighting
│       │   │   ├── depth
│       │   │   └── rgb
│       │   ├── 02_sunlight_window
│       │   │   ├── depth
│       │   │   └── rgb
│       │   ├── 03_floor_backgrounds
│       │   │   ├── depth
│       │   │   └── rgb
│       │   └── 04_robot_scene
│       │       ├── depth
│       │       └── rgb
│       └── full
│           ├── depth
│           └── rgb
└── Synthetic_Train_Sets
    ├── 4k_Material_Randomized
    │   ├── coco
    │   └── yolo
    │       ├── images
    │       │   ├── train
    │       │   └── val
    │       └── labels
    │           ├── train
    │           └── val
    └── 4K_Physics_Intrinsics_RGB_Exp
        ├── coco
        └── yolo
            ├── images
            │   ├── train
            │   └── val
            └── labels
                ├── train
                └── val

Description of Key Folders

Assets

Contains resources for synthetic data generation and running the pipeline

• CADs: 3D models of all 32 parts generated via our four methods: Manual (expert moddeling), 3DGS (3D Gaussian Splattin), MeshyAI (texture generation), and TRELLIS (GenAI 3D asset).
• General: Backgrounds, distractor objects, and plane materials for scene composition.
• 3D_GenAI_Masked_Imgs: Real object images with segmentation masks for GenAI tools.

Comparison between manually modeled synthetic assets (left) and real-world objects (right).

Real_Test_Set

Captured with a Zivid 2 Plus MR60 industrial RGB-D camera.

• annotations/: COCO and YOLO bounding-box annotations.
• images/: RGB images and depth data.

The real test set is provided in two complementary formats: a full evaluation set (images/full/ and annotations/full/) for comprehensive benchmarking across all 508 images, and per-scene organization (images/by_scene/ and annotations/by_scene/) organized into 4 distinct industrial scenarios (controlled lighting, window sunlight, background diversity, and robot-mounted views). This dual structure allows researchers to either evaluate overall performance or conduct targeted analysis of specific environmental challenges.

Synthetic_Train_Sets

Images and bounding box annotations of our two best performing configuration synthetic datasets (4000 images each):

• 4k_Material_Randomized: Manual modelled CADs with material randomization
• 4K_Physics_Intrinsics_RGB_Exp: Manual modelled CADs and textures

Manual modelled CADs and textures (left) and randomized materials (right).

Checkpoints

Pre-trained YOLO11m models of our best 2 performing synthetic datasets:

• yolo11m_Material_Randomized.pt: Trained on 4k_Material_Randomized dataset
• yolo11m_Physics_Intrinsics_RGB_Exp.pt: Trained on 4K_Physics_Intrinsics_RGB_Exp dataset

Object Classes

Prefix Meaning C Custom-Modeled GF Global Fastener MM McMaster F Fath24

Suffix Meaning S Small M Medium L Large

Family / Source	Object/Class Name(s)
Custom-Modeled	C_O_Ring_L, C_O_Ring_M, C_O_Ring_S
	C_Plastic_Washer_L, C_Plastic_Washer_S
	C_Steel_Ball_L, C_Steel_Ball_S
	C_Washer_M5
	C_Washer_M6
FATH GmbH	F_Roll-in_Nut_M5
Festo SE & Co. KG	FestoI
	FestoT
	Festo_Torch
	FestoV
	FestoX
	FestoY
GlobalFastener Inc.	GF_Collar_L, GF_Collar_S
	GF_Slotted_Pin_L, GF_Slotted_Pin_S
	GF_Split_Pin_L, GF_Split_Pin_S
	GF_Cone_Screw_M8
	GF_Hexagon_Nut
	GF_Knurled_Screw_M8
	GF_Plain_Screw_M8
	GF_Screw_M5
McMaster-Carr Supply Co.	MM_Silencer_L, MM_Silencer_S
	MM_Spring
	MM_Wing
	MM_Wood_Screw

Citation

J. M. Araya-Martinez, T. Tom, A. S. Reig, P. R. Valiente, J. Lambrecht, and J. Krüger, “Synthrender and iris: Open-source framework and dataset for bidirectional sim-real transfer in industrial object perception,” 2026. [Online]. Available: https://arxiv.org/abs/2602.21141

License

See LICENSE.txt for terms and conditions.

Contact

For questions, please contact the corresponding authors of the paper.