Classification — Leaned vs Straight vs Rejected
- This dataset contains real-world images of electric utility poles labeled into three classes: "Leaned", "Straight", and "Rejected". It is designed for image-level classification tasks.
- The dataset complements a previously released object detection dataset focused on bounding-box annotations.
- This module fine-tunes a DINOv2 ViT-B/14 model for image-level classification of poles.
- GitHub repo link: Here
Dataset Details
Dataset Description
- Task: Image classification (multi-class)
- Classes:
Leaned,Straight,Rejected - Image format:
.jpg - Label format: Folder-based (one folder per class)
The images were collected by field linemen across three districts in Andhra Pradesh: Visakhapatnam, Eluru, and Kakinada. Each image was labeled by three annotators, with the final label selected via majority vote.
- Curated by: Sampath Balaji & team, at APEPDCL
- Funded by: Eastern Power Distribution Company of Andhra Pradesh Limited (APEPDCL)
- License: CC BY 4.0
- Language(s): Not applicable
Uses
Direct Use
This dataset is intended for training and evaluating machine learning models that classify pole alignment from real-world utility inspection photos.
Out-of-Scope Use
- Surveillance, identity detection, or person-tracking
- Facial recognition or biometrics
- Applications requiring highly controlled camera perspectives
Dataset Structure
The dataset is organized into pre-split folders:
Poles_LeanedStraight/Classification/
├── train/
│ ├── Leaned/
│ ├── Straight/
│ └── Rejected/
├── val/
│ ├── Leaned/
│ ├── Straight/
│ └── Rejected/
├── test/
│ ├── Leaned/
│ ├── Straight/
│ └── Rejected/
Rejectedindicates poles/images that were ambiguous or pole bottom or top was not visible or unusable for downstream tasks but still part of this classification problem.- No augmentations were applied.
- Each split is approximately: Train 75%, Val 15%, Test 10%
Dataset Creation
Curation Rationale
This dataset was created to build an ML pipeline for real-time defect detection in utility poles, starting with identifying leaned poles as a foundational task. The aim is to build an electrical line quality monitoring system to revolutionize maintenance workflows.
Source Data
Data Collection and Processing
Images were collected using mobile cameras by field linemen in Andhra Pradesh.
Image Capture Protocol
The images were captured on the ground by APEPDCL linemen following a field protocol designed to reduce bias and ensure consistency:
- Linemen were instructed to hold the phone upright and parallel to the pole to avoid tilt or skew that could distort the pole's perceived angle.
- Images were to include both the top and bottom of the pole, as full visibility is essential for lean classification.
- No filters or enhancements were applied during capture.
Despite this protocol, many images were captured with a tilted horizon, making it difficult to determine if a pole was leaned or not. Tools like Google Photos and Adobe Lightroom were used to manually correct tilt by straightening the horizon.
To prevent this in future data, the team identified an open-source camera app, Open Camera, which uses the phone’s gyroscope sensor to auto-straighten images. This app is now recommended for field use to ensure more consistent alignment.
Observations and Future Protocol Improvements
During annotation, a key insight emerged:
- When images were taken too close to the pole, perspective distortion made it harder to determine pole alignment.
- Going forward, linemen will be instructed to stand approximately one pole height away to reduce distortion and improve labeling accuracy.
Who are the source data producers?
Electric linemen from APEPDCL working in Visakhapatnam, Eluru, and Kakinada districts of Andhra Pradesh, India.
Annotations
Labeling Process
Images were individually labeled by 3 human labelers and saved in folder-based structure. Majority vote was used to determine the final class label. The team worked to maintain consistency, though some labeling errors may persist due to the manual nature of the task.
Who are the annotators?
The annotation team consisted of project members:
- Srija Vanapalli
- Krishna Sumwith
- Jyotsna Adari
Under the direction of the team lead.
Personal and Sensitive Information
No personal or sensitive data was intentionally included. However, some images may contain people or faces captured incidentally in public areas, as the data was collected in outdoor settings.
Bias, Risks, and Limitations
- Geographic bias: All data comes from Andhra Pradesh, India — results may not generalize to other regions or environments.
- Scene/perspective bias: Variation in camera angles and distances may affect model training.
- Privacy risk: Images taken in public areas may contain identifiable people.
Recommendations
- Apply image normalization or augmentation techniques to improve generalization.
- Account for perspective and camera angle effects during training and evaluation.
- Use ethically and responsibly, especially in systems that affect public infrastructure.
Validation Accuracy
- Best Val Acc: 84.14%
- Hardware used: RTX 4070 Ti SUPER on [JOHNAIC](https://von-neumann.ai)
Sample Visualizations
More Information
Follow-up datasets in this series will include other electrical line and pole defects as part of APEPDCL's Line Quality monitoring project.