From Semantic To Instance: A Semi-Self-Supervised Learning Approach (2506.16563v1)

Abstract: Instance segmentation is essential for applications such as automated monitoring of plant health, growth, and yield. However, extensive effort is required to create large-scale datasets with pixel-level annotations of each object instance for developing instance segmentation models that restrict the use of deep learning in these areas. This challenge is more significant in images with densely packed, self-occluded objects, which are common in agriculture. To address this challenge, we propose a semi-self-supervised learning approach that requires minimal manual annotation to develop a high-performing instance segmentation model. We design GLMask, an image-mask representation for the model to focus on shape, texture, and pattern while minimizing its dependence on color features. We develop a pipeline to generate semantic segmentation and then transform it into instance-level segmentation. The proposed approach substantially outperforms the conventional instance segmentation models, establishing a state-of-the-art wheat head instance segmentation model with mAP@50 of 98.5%. Additionally, we assessed the proposed methodology on the general-purpose Microsoft COCO dataset, achieving a significant performance improvement of over 12.6% mAP@50. This highlights that the utility of our proposed approach extends beyond precision agriculture and applies to other domains, specifically those with similar data characteristics.

• The paper presents a novel semi-self-supervised approach that uses limited human annotations and synthetic data to achieve high segmentation accuracy.
• It leverages a unique three-channel GLMask encoding that decouples color dependency, boosting mAP from ~50% to over 97% in wheat head segmentation.
• The method generalizes well across domains, outperforming standard RGB pipelines on COCO with improvements up to 18% in mAP.

Semi-Self-Supervised Instance Segmentation via GLMask and Synthetic Data

This work introduces a semi-self-supervised learning methodology for instance segmentation in domains where densely packed, self-occluded, and visually homogeneous objects make annotation labor-intensive and standard supervised learning approaches impractical. The methodology synthesizes instance-level data from a minimal set of human-annotated images and leverages a novel image-mask representation, GLMask, to maximize generalization and segmentation performance while minimizing reliance on color information. The approach is evaluated extensively on precision agriculture imagery (notably wheat head segmentation) and general-purpose data (MS COCO), demonstrating substantial improvements over standard RGB-based pipelines in both domain-specific and cross-domain settings.

Methodological Overview

The central contributions and technical advances of the paper are as follows:

• Data Synthesis Pipeline: Utilizing only a small number (10-36) of manually annotated RGB frames, the authors generate large-scale synthetic datasets appropriate for instance segmentation by a modified cut-and-paste approach. This pipeline overlays annotated foreground instances (wheat heads) on diverse, real background scenes, using precomputed instance masks to preserve segmentation fidelity. Variation is introduced via spatial (flip, rotation, elastic transform) and pixel-level (color jitter, channel dropout, blur, noise, etc.) transformations, targeting domain generalization.
• GLMask Encoding: A new three-channel image representation is proposed, defined as the concatenation of grayscale (G), LAB-lightness channel (L), and a semantic segmentation mask (M). GLMask is designed to accentuate shape and textural features while suppressing color-specific cues, addressing generalization under varying phenological (crop maturation, lighting conditions) and acquisition (sensor, environment) variations. This is especially impactful for agricultural data, where color can be highly variable and is frequently confounded.
• Learning Protocol: The core instance segmentation model is based on YOLOv9 in segmentation mode (YOLOv9e-Seg), chosen for its balance between accuracy and inference efficiency. Training proceeds in two major stages:
  1. Pretraining on synthetic GLMask-encoded data, yielding the SynModel.
  2. Domain adaptation via either rotation-augmented real images (RoAModel) or pseudo-labeling of unlabeled real images with the pretrained SynModel as teacher (PseModel). This protocol is paralleled for both the in-domain (wheat heads) and out-of-domain (COCO) settings.

Empirical Results

Evaluation on challenging wheat head datasets reveals:

• Strong Gains over RGB Baseline: The SynModel, trained exclusively on GLMask-encoded synthetic data, attains mAP@50 of 97.9% and 88.2% on challenging, multi-domain wheat head test sets, compared to 50.4% and 49.7% for the equivalent RGB-trained baseline. These gains reflect the value of both GLMask encoding and large-scale data synthesis.

• Further Performance with Domain Adaptation: RoAModel, fine-tuned with rotation-augmented real data, achieves up to 98.5% mAP@50 and 85.3% mAP@50-95, with visually consistent instance outlines (confirmed by cross-domain and qualitative visualizations). Pseudo-label adaptation underperforms relative to rotation augmentation, highlighting the continued necessity of augmenting label diversity in the agricultural context.
• Generalization Beyond Agriculture: On MS COCO 2017, the GLMask-encoded model exceeds the RGB baseline by 12.6% mAP@50 and 17.8% mAP@50-95, even though the GLMask in this context is limited to using binary (not semantic) masks. When only binary masks are available, the technique still confers a robust benefit, which could be further improved with full semantic mask access.

Discussion and Implementation Considerations

Practical Implications:

• Data Efficiency: The synthesis pipeline eliminates the need for manual annotation at scale, reducing data collection and labeling bottlenecks, especially in domains with many occlusions and repeating small objects (e.g., agriculture, medical imaging, materials science).
• Model Generalization: By intentionally decoupling segmentation from color bias through the GLMask, trained models show robustness to environmental and phenotypic variability, facilitating application in real-world field settings with uncontrolled illumination and crop diversity.
• Computational Requirements: The primary compute load arises in data synthesis and large-scale YOLO training. The experiments are conducted on modern multi-GPU infrastructure (A40, V100S), with typical batch sizes 18 (for 1024x1024 wheat images) up to 48 (for 640x640 COCO images). The approach is therefore suitable for production settings where ample compute is available, but real-time inference is achievable due to efficient YOLO variants.

Limitations:

• Fragmented/Occluded Object Failures: The methodology occasionally mislabels fragments of occluded objects as separate instances, and merges overlapping instances, particularly in highly crowded or high-altitude fields. Such errors are not unique to this method but are highlighted as failure points warranting further architectural investigation.
• Dependency on Segmentation Quality for GLMask: The technique presumes access to an adequate semantic segmentation model (potentially trained via self-supervised or efficient weakly-supervised means), especially in the cross-domain setting.

Future Opportunities:

• Augmentation Beyond Rotation: Exploring richer forms of geometric and photometric augmentation, or leveraging unsupervised domain adaptation with adversarial training, could enhance domain generalization further.
• Broader Architecture Applicability: Extending GLMask input encoding to foundation segmentation architectures (e.g., SAM/Mask2Former) could validate the technique's generality and potentially further boost instance-wise performance.
• Multi-Class & Hierarchical Segmentation: For multi-class or hierarchical instance segmentation (common in medical and environmental imagery), developing semantic mask encodings beyond the binary channel used in the present COCO experiments could provide further gains.

Conclusion

The paper establishes that semi-self-supervised training strategies, guided by minimal manual annotation, targeted data synthesis, and representation engineering via GLMask, can yield SOTA instance segmentation performance on both domain-specific and general-purpose tasks. The GLMask input encoding, in particular, offers a practical and readily reproducible means to mitigate model over-reliance on color, benefiting diverse application areas featuring repetitive and occluded objects. The approach is a significant step towards scalable, annotation-light deep learning for structured segmentation problems, especially under data or resource constraints.