A Normalized Gaussian Wasserstein Distance for Tiny Object Detection

Source: Dev.to

Overview

Finding tiny objects in images is challenging because they occupy only a few pixels and can be missed by methods that expect larger shapes. Traditional evaluation metrics based on overlap are not well suited for such small targets.

Method

Researchers propose a new way to compare predicted bounding boxes with ground‑truth boxes. Instead of using overlap, each box is treated as a soft, blurry spot and represented as a 2D Gaussian. The similarity between the predicted and ground‑truth boxes is then measured using the Normalized Wasserstein Distance (NWD).

• NWD is less sensitive to tiny shifts, allowing detectors to recognize objects even when they move slightly.
• It reduces false negatives by providing a more tolerant metric for small objects.

Experiments

The NWD metric was integrated into modern object detectors and evaluated on the AI‑TOD dataset, which is specifically designed for tiny object detection. Results show a significant improvement over traditional overlap‑based metrics, especially for tiny objects.

Applications

Because the change is simple to implement and does not require substantial additional computation, it can be adopted in devices such as smartphones and cameras. This enhances the ability to detect small items like drones, birds, or distant cars, making vision systems more reliable in scenarios where fine details matter.

Read the full paper:
A Normalized Gaussian Wasserstein Distance for Tiny Object Detection