Course contentsShow
← Machine Learning and Deep Learning
Lesson 982 of 3,53822. Computer Vision TasksFree lesson

Atrous Spatial Pyramid Pooling (ASPP)

Learn how ASPP captures multi-scale context by applying parallel atrous convolutions at different dilation rates.

Atrous Spatial Pyramid Pooling (ASPP)

What you'll learn: How ASPP captures information at multiple scales simultaneously by running parallel atrous convolutions with different dilation rates.

The Core Idea

Remember atrous (dilated) convolutions from our previous lessons? They expand the receptive field without adding parameters by inserting gaps between kernel weights. ASPP takes this concept further by applying multiple atrous convolutions in parallel, each with a different dilation rate, then combining their outputs.

Think of ASPP as looking at an image through multiple magnifying glasses at once—some show fine details (small dilation), others reveal broader context (large dilation). This multi-scale view helps segmentation networks understand objects at various sizes without needing to resize the input image multiple times.

How ASPP Works

A typical ASPP module contains several parallel branches:

  • One 1×1 convolution (captures point-wise information)
  • Three or more 3×3 atrous convolutions with different dilation rates (e.g., rates 6, 12, 18)
  • Often a global average pooling branch (captures image-level context)

All branches process the same input feature map simultaneously. Their outputs are then concatenated and passed through another 1×1 convolution to blend the multi-scale features together.

This parallel structure is why it's called a "pyramid"—you're sampling spatial information at multiple scales (rates), similar to how spatial pyramid pooling works but using atrous convolutions instead.

Why It Matters for Segmentation

In semantic segmentation, objects can appear at vastly different scales. ASPP lets the network see both small details and large contextual regions without sacrificing resolution, making it especially powerful in DeepLab architectures.

Key Takeaway: ASPP captures multi-scale context efficiently by running parallel atrous convolutions with different dilation rates and combining their outputs, giving segmentation networks a richer understanding of objects at various scales.