A High Precision Annotation Tool for Computer Vision and Research Workflows

Intuitus is pleased to announce Lepakshi Annotator, a Windows-native, on-premises annotation desktop platform built from the ground up using Microsoft Foundation Classes (MFC). Engineered specifically for high-consequence domains like agricultural phenotyping and botanical analysis, Lepakshi delivers fluid, pixel-accurate anatomical contouring without slowing annotation workflows.

Dual Workflow: Precision Manual Crafting Meets AI Assistance

Lepakshi Annotator is designed to handle multi-megapixel leaf images with zero canvas lag and low memory overhead, offering annotators two complementary routes:  

• Full-Fidelity Manual Segmentation: An ergonomic, sub-pixel manual workflow for plotting, editing, and smoothing irregular biological contours directly on the canvas.  

• Pre-Trained YOLO-Seg Inference Integration: When an existing, previously trained YOLO-Seg instance-segmentation model is available, annotators can invoke it on demand. Rather than drawing from scratch, the deep learning model automatically predicts candidate object silhouettes, providing instant polygonal proposals [1,2] that users can review, reshape, and approve in seconds.

Why On‑Premise Annotation Matters

Cloud-based platforms offer convenience, scalability and collaboration, but third-party hosting may not suit sensitive or proprietary datasets. Agriculture, biomedical signals, industrial inspection images and government-regulated data can require tighter control over storage and access. For these workflows, local deployment provides an alternative to uploading raw images to an external service.

On-premises annotation supports the following data-handling and offline-workflow goals; achieving them depends on deployment configuration and internal controls:

• Data never leaves the local machine or local network.

• No external storage, cloud syncing, or vendor access

• Full compliance with internal security policies

• Protection of intellectual property and confidential datasets

• Deterministic, offline operation without dependency on internet connectivity

The Global Annotation Software Market: Growth and Opportunity

The AI annotation market has evolved from a niche service into a foundational industry. According to current projections, the market stood at USD 2.3 billion in 2024 and is expected to grow at a CAGR of 28.6%, reaching approximately USD 28.5 billion by 2034 [3]. This explosive growth is driven by the rising need for high‑quality training data across sectors such as healthcare, agriculture, autonomous systems, industrial inspection, and public‑sector analytics.

Why Annotation Tools Are Critical—and Where Intuitus Fits

Intuitus’s opportunity lies in supporting domain-specific AI development through precise annotations, human validation and controlled, reproducible data workflows.

Lepakshi’s agricultural application illustrates how these priorities come together in practice.

• High-precision vision models: Pixel-accurate masks make annotation precision a key priority for Intuitus.

• Human validation: Combining AI-assisted labelling with human refinement provides a clear direction for Intuitus.

• Compliance requirements: Secure, on-premise workflows offer a strong positioning offer from Intuitus when data control and reproducibility matter.

Lepakshi Annotator

Agricultural phenotyping and botanical analysis require accurate leaf boundaries without slowing annotation. Multi-megapixel leaf images make responsive navigation and practical polygon editing especially important. Lepakshi’s dedicated C++/MFC Windows client addresses these requirements by combining manual curation with neural-network inference.

The Windows client handles image navigation and polygon editing, while a separate local deep-learning service runs in WSL2, the Windows Subsystem for Linux. This separation supports a human-in-the-loop workflow in which users review AI-generated segmentation proposals and refine irregular regions of interest.

Features

• Native C++/MFC client: A Windows desktop architecture designed for responsive rendering of multi-megapixel scientific images with low memory overhead, without relying on a web framework.

  

  Figure 1.0 Manual segmentation interface displaying bounding box drawing capabilities by placing seed points on a leaf, illustrating the process of defining a region of interest.

• Bounding boxes and polygons: Annotate objects with axis-aligned boxes or detailed polygons that follow petioles, leaf apices and asymmetric biological contours.

• Real-time curve smoothing: Turn polygon vertices into smooth curves directly on the canvas using Chaikin’s corner-cutting subdivision and Cardinal spline interpolation.

• Interactive boundary editing: Adjust edges, insert vertices and refine contours directly with on-canvas control handles

Figure 2.0. Real-time curve smoothing and seed-point editing workflow. Candidate boundary seed points are generated via a decoupled FastAPI inference service running YOLO-Seg in WSL2, overlaid on the native MFC canvas for interactive polygon refinement.

• Asynchronous inference integration: Send requests and receive predictions through REST interface. The client connects to a separate deep-learning service, such as a local FastAPI service running PyTorch in WSL2.

• AI-assisted pre-annotation: Overlay segmentation proposals from YOLO-seg or specialised edge models, then verify, clean up and convert the masks into polygons. Refer Figure 3.0 [1,2]

References

[1] D. Bolya, C. Zhou, F. Xiao, and Y. J. Lee, "YOLACT: Real-time instance segmentation," in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Seoul, Korea (South), 2019, pp. 9157–9166, doi: 10.1109/ICCV.2019.00925.

[2] Jocher, G., Qiu, J., & Chaurasia, A. (2023/2024). Ultralytics YOLO [Computer software]. https://github.com/ultralytics/ultralytics

[3]  AI Annotation Market Size | CAGR of 28.60%