Multi-Stage Residual Refinement for Apple Segmentation on MinneApple

Keonvin Park, Jin Hong Mok

Keywords: Apple segmentation, Multi-stage residual refinement, MinneApple dataset, Agricultural computer vision, Accuracy-efficiency trade-off

Abstract:

Accurate and efficient apple segmentation is critical for automated agricultural analysis and yield estimation. The MinneApple dataset has emerged as a standard benchmark for evaluating detection and segmentation performance in orchard environments. While modern convolutional architectures achieve competitive mean Intersection-over-Union (mIoU), the impact of progressive refinement depth on segmentation accuracy and computational efficiency remains underexplored.
In this work, we investigate multi-stage residual refinement for apple segmentation on the MinneApple dataset. Starting from a strong UNet-ResNet baseline (1-stage), we introduce a progressive residual correction framework in which each additional stage predicts a residual mask to refine the previous output. Importantly, we increase refinement depth without modifying the backbone, enabling controlled analysis of refinement behavior.
Through a systematic study from one to five refinement stages under identical training settings, we observe consistent accuracy gains from 0.6482 mIoU (1-stage) to 0.6904 (3-stage), and up to 0.6971 with five stages. However, improvements are not strictly monotonic, with intermediate saturation observed at four stages. Notably, inference speed remains largely stable across stages (approximately 235–273 FPS in our implementation), indicating that lightweight residual refinement can improve segmentation accuracy without substantial runtime degradation.
These results demonstrate that progressive residual correction effectively enhances apple segmentation on MinneApple, while revealing diminishing returns beyond moderate refinement depth. Our findings provide practical guidance for designing efficient multi-stage segmentation systems in agricultural vision applications.

Submission Number: 8