Swara's Weekly Update: Building a Dataset for the National Loon Center

A machine learning model is only as good as the data it learns from, and nowhere is that clearer than in a project trying to teach a computer to spot loons in drone footage. Swara Joshi's second video in her computer vision series for the National Loon Center project steps back from the model itself to focus on the part that actually determines whether the model will work: the dataset.

From raw images to structured information

A dataset, in this context, is not just a folder of photos. It is a collection of images paired with information about what objects are in each one and exactly where those objects sit. Handing a model a raw image of a loon tells it nothing useful on its own; the model also needs to be told where in that image the loon actually is. That is where annotation comes in. Using a tool called CVAT, Swara uploads an image, draws a bounding box around the loon, and assigns it a label. The image goes from an undifferentiated photo to one carrying structured information the model can actually learn from.

Consistency is harder than it sounds

Drawing a box around a bird sounds simple until you start asking the follow-up questions. How tightly should the box fit the loon's body? What happens when only part of the loon is visible, obscured by water or another bird? What about a loon that appears very small in the frame, or a frame containing several birds at once? Every one of these situations becomes an example the model will eventually learn from, which means every inconsistent answer to these questions becomes a small distortion in what the model thinks a loon looks like.

Why variation matters as much as volume

It would be easy to assume that more images automatically means a better dataset, but the video makes a sharper point: if every training image shows a loon from the same distance, the same angle, and the same lighting, the model risks learning those specific conditions rather than learning what a loon actually looks like in general. Deliberately building in variety, different distances, angles, lighting conditions, water backgrounds, and positions, forces the model to generalize instead of memorizing a narrow set of circumstances.

Starting small on purpose

Rather than annotating a huge volume of images immediately, the project is deliberately building a small, consistent pilot dataset first. The goal at this stage is establishing the foundation and refining the annotation guidelines, not producing a final detector. Learning from the annotation process on a manageable pilot set, and improving the guidelines based on what that process reveals, is treated as a necessary step before scaling up once more drone footage becomes available.

The full pipeline, start to finish

Zooming out, the overall workflow moves through a clear sequence: collect images, prepare them, annotate them, review the annotations, build the training dataset, and then train the model. Once enough annotated images exist, they get exported into a format such as YOLO, which can then train an actual detection model. The video's central message is that this entire pipeline starts with data, not with model architecture. Good quality, consistently labeled, and appropriately varied data is what gives a model like this a real chance of working on unpredictable, real-world drone footage.

Key takeaways

• A dataset pairs images with structured information about what objects appear in them and exactly where, which requires annotation, not just raw photos.
• CVAT is used to draw bounding boxes around loons and assign labels, turning raw images into training-ready data.
• Consistency questions, box tightness, partial visibility, small objects, and multiple birds per frame, directly affect what the model learns.
• Deliberate variation in distance, angle, lighting, and background prevents the model from overfitting to narrow conditions.
• The project is starting with a small, consistent pilot dataset to refine annotation guidelines before scaling up.
• The full pipeline runs collect, prepare, annotate, review, build the dataset, then train, with a YOLO model as the eventual target.

Who this is for

This is for anyone learning how computer vision datasets are actually built, students following the National Loon Center project as part of the Humanitarians AI Fellows program, and conservation-minded developers interested in applying object detection to wildlife monitoring from drone footage.