Detection and Segmentation of Custom Objects using High Distraction Photorealistic Synthetic Data

TL;DR

This paper presents a methodology for instance segmentation of custom objects using photorealistic synthetic data, demonstrating its effectiveness through a new dataset and analysis of robustness against distractions.

Contribution

The authors introduce a novel synthetic dataset and a domain randomization technique for training instance segmentation models on custom objects, reducing reliance on manual data collection.

Findings

Synthetic data achieves high segmentation accuracy on real-world images.

Distraction objects improve model robustness to occlusion and lighting variations.

The dataset and methodology facilitate scalable training for custom object detection.

Abstract

We show a straightforward and useful methodology for performing instance segmentation using synthetic data. We apply this methodology on a basic case and derived insights through quantitative analysis. We created a new public dataset: The Expo Markers Dataset intended for detection and segmentation tasks. This dataset contains 5,000 synthetic photorealistic images with their corresponding pixel-perfect segmentation ground truth. The goal is to achieve high performance on manually-gathered and annotated real-world data of custom objects. We do that by creating 3D models of the target objects and other possible distraction objects and place them within a simulated environment. Expo Markers were chosen for this task, fitting our requirements of a custom object due to the exact texture, size and 3D shape. An additional advantage is the availability of this object in offices around the world for easy testing and validation of our results. We generate the data using a domain randomization technique that also simulates other photorealistic objects in the scene, known as distraction objects. These objects provide visual complexity, occlusions, and lighting challenges to help our model gain robustness in training. We are also releasing our manually-gathered datasets used for comparison and evaluation of our synthetic dataset. This white-paper provides strong evidence that photorealistic simulated data can be used in practical real world applications as a more scalable and flexible solution than manually-captured data. Code is available at the following address: https://github.com/DataGenResearchTeam/expo_markers