Overview
Smart Inspections with Vision was a CJ Logistics / CJ TES object-detection project for logistics-cart inspection. The task was to detect many product classes in cart images even though the available training set mostly contained isolated object photos rather than full cart scenes.
Challenge
- Evaluation images were large cart images, commonly around 3023×4032 or 3023×3023 pixels.
- The target set contained 100 classes with multiple instances per class.
- Some objects appeared from difficult angles or were partially occluded.
- Real evaluation scenes also included unknown objects outside the target class list.
Approach
- Added validation-style masks into the training workflow after manual annotation with CVAT.
- Generated synthetic training scenes by placing scalable object crops onto grid-plane and cart-like layouts.
- Used cut-and-paste augmentation so isolated product images could resemble warehouse/cart scenes.
- Added selected unknown-object data to improve robustness against distractors.
- Trained YOLOv8-X with PyTorch, OpenCV, and Albumentations.
Issues addressed
Random object placement often produced overlapping bounding boxes, so placement was filtered by IoU constraints. Blur was added to reflect focus variation, while lighting flare remained a known gap for future synthetic-data improvement.
Result
The final solution used the cut-and-paste synthetic-data strategy and trained YOLOv8-X for 20 epochs. It reached fourth place in the CJ Logistics / CJ TES Future Technology Challenge evaluation, with IoU 0.72 and accuracy 0.264 in the recorded result.
Materials
This visual summarizes the logistics-cart inspection pipeline: synthetic data generation, object detection, and evaluation under difficult cart-scene viewpoints.
