In modern warehouses, where missort rates of 20-30% translate into millions in annual losses, even small improvements in picking accuracy can generate significant savings. Traditional solutions often struggle with damaged or non-standard boxes and packages, creating a persistent efficiency gap in automation systems.

‍

Role of Synthetic Data Solution

We started with 15 images, representing variations in background, environmental changes, and box size and quantity. We used the Advex platform to create 60 auto-labeled images with segmentation masks. These images covered a wide range of edge cases, with variations in textures, lighting, stacking orientations, and damage profiles.

                                                                                                                                     _{Image: Advex Synthetically Generated Images} 

‍

Performance and Insights: A Stark Lesson in Domain Specificity

We started with just 12 real images and added 60 synthetic examples. This use case produced the most dramatic results of the study.

‍

Winning Model Enhancement: 

Segformer with NVDINOv2 foundation model as backbone thrived in this complex, dynamic environment. It leveraged the synthetic data to achieve a remarkable 39.66% surge in its mIoU score, transforming it from a poor performer into the best model for the task.

‍

                                                                                                                   ^{Graph: NVDINOv2 Performance Metrics for package segmentation by a Robotic Arm} 

‍

‍

Importance of Model Selection: 

Conversely, C-RADIOv2, the star of defect detection, saw its performance drop by 18.31%. This illustrates a classic “domain shift”, where a model may perform well for one type of task, such as defect detection, but a different backbone like NVDINOv2 may perform better for a different type of task, such as box segmentation.

‍

                                                                                                               _{Graph: Box Segmentation Model mIoU Performance Uplift Across 3 Models}

‍

‍

From Metrics to Reality

By enriching the training set with diverse synthetic data, the model learns to segment brown boxes with far greater confidence and precision. The following comparison provides a clear visual illustration of this principle in action.

‍

                                                                                                                             _{Image: Before and after comparison of robotic arm box segmentation}

‍