Improving small object detection

WHAT THIS TESTS The interviewer wants to see structured debugging of a known failure mode rather than random hyperparameter twiddling. Small objects are hard because they cover few pixels and get erased as feature maps are downsampled through the backbone.

A GOOD ANSWER COVERS Three distinct levers. First, input and feature resolution: increase the network input size, slice large images into overlapping tiles at inference and training, and detect on higher-resolution feature maps. Second, multi-scale architecture: use a Feature Pyramid Network so shallow, high-resolution layers carry strong semantics, and assign small objects to those finer levels. Third, data and anchor strategy: add anchors or query priors sized for tiny objects, rebalance the loss with focal loss, and use targeted augmentation such as small-object copy-paste, mosaic, and scale jitter so the model sees more small instances.

COMMON WRONG ANSWERS Saying just train longer or use a deeper backbone. Depth alone increases downsampling and can hurt small objects. Another weak answer is only collect more data without changing how scale is handled, since the architecture may still discard the relevant signal.

LIKELY FOLLOW-UPS How does an FPN actually help, and which pyramid level should detect a 12-pixel object. How do you set the IoU threshold and anchor scales for tiny boxes. What metric isolates small-object performance, and the answer is COCO AP small. How does tiling change non-maximum suppression at tile boundaries.

ONE CONCRETE EXAMPLE In aerial imagery of cars, a detector trained at 640 pixels misses vehicles that span only a handful of pixels. Tiling each 4000-pixel image into 640 patches with overlap, switching to an FPN-based detector, and adding copy-paste of cropped cars raises small-object recall substantially, while overlap plus global NMS reconciles boxes at the seams. You then track COCO AP small specifically rather than overall AP, so gains on tiny objects are not masked by easy large ones, and you confirm that the finer pyramid levels are the ones firing on these instances.