ASPP
A multi-scale feature extraction module introduced in the DeepLab series. It uses parallel atrous (dilated) convolutions with different dilation rates to capture contextual information and enlarge the receptive field without reducing feature map resolution.
Timeline (last 90 days)
DeepLabV3 extracts multi-scale features in parallel at different dilation rates through the ASPP structure, effectively addressing the problem of scale variation in segmentation targets.
ASPP (Atrous Spatial Pyramid Pooling) was proposed in the DeepLab series of works and uses multiple parallel atrous convolutions with different dilation rates to extract multi-scale contextual information.
ASPP is typically placed at the end of the encoder (at the deepest feature map), because this is where semantic information is richest and spatial resolution is lowest, making the benefits of multi-scale receptive fields greatest.
The third round of combination experiments involved four directions—directional convolution, deep supervision, EMA, and ASPP—theoretically yielding 16 combinations, with the author stopping at the 13th combination.
All Facts (4)
DeepLabV3 extracts multi-scale features in parallel at different dilation rates through the ASPP structure, effectively addressing the problem of scale variation in segmentation targets.
50%UnverifiedASPP (Atrous Spatial Pyramid Pooling) was proposed in the DeepLab series of works and uses multiple parallel atrous convolutions with different dilation rates to extract multi-scale contextual information.
50%UnverifiedASPP is typically placed at the end of the encoder (at the deepest feature map), because this is where semantic information is richest and spatial resolution is lowest, making the benefits of multi-scale receptive fields greatest.
50%UnverifiedThe third round of combination experiments involved four directions—directional convolution, deep supervision, EMA, and ASPP—theoretically yielding 16 combinations, with the author stopping at the 13th combination.
50%