Computer vision
YOLO Object Detection in Practice
Built a reusable object detection pipeline from annotation and augmentation to training, tuning and evaluation, with result visualisation.
Background and problem
I needed a reusable object detection pipeline that covers the full path from raw images to an evaluable model, so that every experiment did not require writing scripts by hand again.
Constraints
- Solo work over about 4 months.
- Compute was limited, so training and tuning had to be completed on a single GPU.
- The data came from public scene images and had to be annotated and cleaned myself.
Approach and architecture
The pipeline has four stages: data annotation and splitting, augmentation and preprocessing, training and tuning, and evaluation and visualisation. Scripts are split by stage and configuration is centralised in YAML to keep experiments reproducible.
Key decisions
- Annotation format: use the official YOLO format to reduce conversion steps.
- Augmentation strategy: keep only the geometric and colour augmentations that are effective for these scenes, avoiding overfitting.
- Evaluation metrics: use mAP@0.5 as the primary metric with inference latency as a secondary one, balancing accuracy and usability.
Results and measurement
Completed more than 40 training rounds and produced reusable training and evaluation scripts; detection visualisations were used for manual review.
Ran the complete loop from annotation to training to evaluation, and the training and evaluation scripts can be reused for similar scenarios; specific mAP and latency figures are subject to the experiment records.
Technical details
- Standardised the data into the YOLO annotation format and fixed the training/validation split.
- Centralised training, augmentation and evaluation parameters in YAML to reduce script duplication.
- Evaluated with mAP@0.5 together with inference latency, and visualised results for review.
Reflection
Data quality affects the final result more than model architecture does. If I did it again, I would establish annotation guidelines and a spot-check mechanism before starting training.
Scope and security
This is a static case study. The page does not run project code or collect uploads, logins or visitor data. Metrics come from project records and local experiments; reproduce details from the source or report. Read the site security model →
