Conventional CNN architectures require fixed input dimensions, often necessitating cropping or warping. These transformations can distort objects or cut off essential features, potentially reducing model accuracy. To solve this, He et al. introduced SPP-Net in 2014, the first model to integrate Spatial Pyramid Pooling into a CNN.

Unlike standard CNNs where fully-connected layers require a fixed-size tensor, SPP-Net uses an SPP layer between the convolution and fully-connected layers. This layer divides the spatial dimension into multiple grids—typically 4x4, 2x2, and 1x1—and performs max-pooling within each. This process produces a fixed-length vector regardless of the input image resolution.

By using this mechanism, SPP-Net achieves robustness against overfitting. It preserves spatial information through smaller pooling windows while capturing a holistic view through larger ones. This results in invariance to object deformation and spatial layout, allowing the model to recognize objects even if they are shifted or slightly deformed.