Preparing Data for YOLO: Training and Evaluating Object Detection Models | Swara Joshi

Swara Joshi continues a series on building a computer vision pipeline for detecting and counting loons from drone footage, moving from labeling images to actually training a YOLO object detection model. The video explains what happens between having annotated images and having a model that can find a loon in a new frame, and why the quality of that in-between step decides whether the model works at all.

From labeled images to a trained model

The pipeline starts with images where loons have already been marked and labeled. YOLO, short for "you only look once," is an object detection model built to scan an image efficiently and return not just whether an object is present but where it is, drawing a bounding box around each loon and attaching a confidence score, such as "loon, 92 percent confidence" at a specific location in the frame. Getting there requires converting the raw annotations into a format the model can actually train on, an intermediate step that sits between labeling and training and is easy to underestimate.

Why annotation quality decides model quality

The workflow runs from images to annotations to dataset preparation to model training to model evaluation to detection. If the annotations feeding that pipeline are inconsistent, incomplete, or inaccurate, the model learns those same mistakes. Joshi frames this directly: the annotation and dataset stages were never separate from the machine learning step, they are the foundation it depends on. Quality upstream produces quality downstream, and there is no later step that fixes a poorly labeled dataset.

Testing on conditions the model has not seen

Once trained, the model is evaluated on images it did not see during training, measuring missed detections, incorrect detections, and how accurately each bounding box locates a loon. Because this is a conservation project, that accuracy matters in a concrete way. A model that scores well on one batch of images can fail when conditions change: drone altitude, lighting, water reflections, camera angle, weather, and distance to the birds all vary in real footage. That is why the dataset needs to grow with diverse examples across those conditions rather than simply accumulating more images that look like the ones already collected.

The honest state of the project

Joshi is explicit that the project does not yet have a production-ready loon detection model and is not claiming one. The current goal is establishing the full pipeline, from raw footage through metadata, annotation, training data, and the model itself, feeding into a researcher-facing application where a researcher can upload footage, run an analysis, and review results without needing to know YOLO is running underneath. The next stage of the series will go deeper into how YOLO training itself works, including validation, predictions, and confidence scores.

Key takeaways

• YOLO detects objects by drawing a bounding box and assigning a confidence score, rather than just flagging that an object is present somewhere in the frame.
• Converting raw annotations into model-ready training data is a distinct, necessary step between labeling and training.
• Inconsistent or inaccurate annotations directly degrade the trained model, since the model learns from whatever it is given.
• Evaluation on unseen images measures missed detections, false detections, and localization accuracy, not just overall presence or absence.
• Real-world variation in altitude, lighting, reflections, weather, and camera angle means the training dataset must include diverse conditions to generalize.

Who this is for

Anyone following a conservation computer vision project from the ground up, or a student wanting a concrete walkthrough of what stands between an annotated image dataset and a working object detector.