AI Vision System - YOLO Dataset Preparation

Title: Image Annotation Specialist (LabelMe) + YOLO Dataset Preparation (34–36 Classes) — Phase 1: Image Scrubbing & Classification Phase 1 Only

Description:

I am building an AI-based visual inspection system using real-world highway images and need support preparing a high-quality dataset for YOLO training.

This is a structured, multi-phase project. Accuracy, consistency, and attention to detail are critical.

⚠️ Important: This job will begin with Phase 1 only (image scrubbing and classification).
Further phases (annotation and YOLO dataset preparation) will follow based on performance.

---

Scope of Work:

Phase 1 – Image Scrubbing & Pre-Classification (Current Phase)

This is the most critical step of the project.

* Review large volumes of real-world highway images (hundreds of thousands available)
* Identify and filter out images that are not useful (no defects, irrelevant content, low-quality data)
* Sort and group images into the correct defect classifications based on provided examples
* Ensure consistency when assigning images to classes (similar defects must always be grouped the same way)

Note:
These are real-world inspection images. Multiple defect types may appear across different stretches of highway, and some images may contain no relevant defects at all. Strong judgment is required. Please message if you have any questions, i need acuracy and strong attention to detail.

---

Phase 2 – Image Annotation (Future Phase)

* Use LabelMe to annotate images
* Draw bounding boxes or polygons depending on object type
* Label objects according to a predefined class list (34–36 classes)
* Follow strict naming and labeling conventions

---

Phase 3 – Dataset Preparation for YOLO (Future Phase)

* Convert LabelMe annotations into YOLO format
* Ensure correct class IDs and structure
* Organize dataset into train/ and val/ folders
* Verify all images have matching label files

---

Phase 4 – Quality Control (Ongoing)

* Ensure labels are accurate and consistent
* Avoid missing or incorrect annotations
* Perform validation before delivery

---

Class System:

* 34–36 defect classes
* Each class will be provided with example images
* All classes are important — the goal is to reflect real-world conditions, not prioritize a subset
* Consistency across similar defect types is critical

---

Requirements:

* Experience reviewing or organizing large image datasets
* Strong attention to detail and consistency
* Ability to follow structured instructions and class definitions
* Familiarity with LabelMe or similar tools is a plus
* Basic understanding of YOLO format is a plus (required for later phases)

---

Deliverables (Phase 1):

* Scrubbed and filtered image sets
* Images grouped into correct classifications
* Clean and organized folder structure

---

Volume:

* Very large dataset (hundreds of thousands of images available)
* Initial batches will be provided for Phase 1
* Only a subset of images will move forward to annotation
* Potential for ongoing work across multiple phases

---

To Apply:

Please include:

* Confirmation that you understand Phase 1 is focused on image scrubbing and classification
* Your approach to reviewing and filtering large image datasets
* Your expected turnaround time for an initial batch
* Any relevant experience with image datasets or annotation work

---

Test Task:

A small test batch will be provided.
You will be asked to scrub and classify images based on provided examples.

---

Notes:

* Accuracy and consistency are more important than speed
* This is part of a larger AI system — data quality is critical
* Strong performance in Phase 1 may lead to continued work in annotation and dataset preparation phases