What is deformable convolution and why is it useful in vision models?

Deformable convolution adapts sampling locations dynamically to better model geometric variations in objects.

Standard convolution samples fixed grid locations (e.g., 3x3 kernel). This rigid structure struggles with objects that have deformations, rotations, or scale changes. Deformable convolution introduces learnable offsets to kernel sampling points, allowing the network to shift receptive fields adaptively based on input features. This improves modeling of complex spatial transformations, especially in object detection and segmentation.