Dilated Convolution (Atrous Convolution)
Dilated Convolution expands the receptive field of a filter by inserting gaps between filter values – larger context without more parameters.
Dilated Convolution inserts gaps in filters for larger receptive fields without extra cost – standard in segmentation and WaveNet.
Explanation
A 3×3 filter with dilation rate r=2 effectively has a 5×5 receptive field but only 9 parameters. By stacking different dilation rates, a network can achieve exponentially growing receptive fields (WaveNet). Standard in semantic segmentation (DeepLab).
Marketing Relevance
Enables global context in dense prediction (segmentation, detection) without downscaling the image.
Origin & History
Yu & Koltun (2015) popularized dilated convolutions for dense prediction. DeepLab (Chen et al., 2017) made atrous convolution standard in semantic segmentation. WaveNet (2016) used causal dilated convolutions for audio generation.
Comparisons & Differences
Dilated Convolution (Atrous Convolution) vs. Standard Convolution
Standard conv: receptive field = filter size; Dilated conv: receptive field = (k-1)·r+1, same parameter count.