Classification networks throw away spatial resolution on purpose — pooling and strided convolutions are how a net trades "where" for "what." Semantic segmentation needs both at once, and for years the field fought this with clumsy fixes: downsample to gather context, then upsample and hope the boundaries survive. This paper's quiet insight was that the resolution loss was never necessary in the first place.
Key Findings
- Receptive field without resolution loss. Inserting zeros between filter taps (dilation) widens what each neuron sees, so context grows exponentially across layers while the feature map keeps its full size — no pooling, no upsampling artifacts.
- No extra parameters or compute per layer. A dilated 3x3 filter still has nine weights; only its reach changes. Aggregating multi-scale context becomes essentially free.
- A drop-in context module. The aggregation block plugs into existing segmentation front-ends and improves dense-prediction accuracy without retraining the whole pipeline from scratch.
- Stripping the classification baggage. The authors also show that the adapted classification net carries vestigial structure that hurts dense prediction, and simplifying it helps.
How It Works
Dilated convolution generalizes the standard kernel with a spacing factor: dilation 1 is ordinary convolution, dilation 2 skips every other pixel, and so on. Stacking layers with exponentially increasing dilation lets the receptive field double each step, covering an entire image region in a logarithmic number of layers — all at native resolution.
Great Fit / When to Skip
Read it if you work on any pixel-dense task — segmentation, depth, or generative models where the dilation idea later resurfaced (WaveNet, DeepLab). Look elsewhere if you want a turnkey modern segmentation recipe: this is the mechanism, not the architecture, and transformer-based segmenters now offer different tradeoffs. Its lasting value is the concept, which quietly became standard equipment.
